OpenClaw and Hermes Agent in November 2026

Every month, TekLoon used to find two utility bills in Gmail, download the attachments, file them in Dropbox, read the amounts and prepare a message requesting payment from his rental operator. He estimates that the routine took about fifteen minutes. His OpenClaw setup now collects the bills and sends him the prepared message through WhatsApp. It is a small improvement, but the task, the documents and the intended result are all identifiable [1].

Henrique Macedo uses Hermes Agent for a different recurring annoyance: booking gym classes when their reservation windows open. Each morning he receives a filtered list on Telegram, replies with the classes he wants, and lets the system book them at the appropriate time. His first attempt consumed about 700,000 tokens every morning. After moving the predictable checks out of the AI model, he reports roughly 6,000 tokens for the same morning exchange [2]. Tokens are the pieces of text and other input that model providers meter; these figures describe his workload, rather than a price promised to every user.

Both agents connect a conversation to applications that perform work. A recurring job has to deliver the correct result on schedule, and its owner has to maintain the connections. The accounts below describe the routines people delegated, the results they received and the repairs they made.

This review covers 13 November 2026. It draws on firsthand accounts from people using the agents at home and at work, alongside official release announcements. The user accounts describe earlier points in 2026 and are dated in the references; they are not retests of the November releases. Reported savings and bills belong to the authors' setups, and EFFOMA has not independently operated those systems or audited their invoices.

OpenClaw's first year and OpenAI funding

Peter Steinberger began the project in November 2025 by connecting messaging to an agent running on his computer. It passed through the names Clawdbot and Moltbot before becoming OpenClaw; the January announcement explains that an Anthropic trademark request prompted the first rename [3]. The enduring idea was straightforward: an assistant available through an ordinary messaging app, with access to selected files and services, that could continue working between conversations.

The frequently repeated claim that OpenAI bought OpenClaw needs correcting. Steinberger announced on 15 February that he was joining OpenAI and that the project would move to an independent foundation [4]. The November project statement identifies OpenAI as a donor, not an owner. The OpenClaw Foundation was introduced in July as an independent nonprofit, with OpenAI supporting work including inference and security [5], [6]. Those sources do not document an acquisition price or a transfer of ownership to OpenAI.

Nous Research released Hermes Agent on 25 February, emphasizing saved knowledge and reusable procedures [7]. A procedure, called a skill in both ecosystems, can preserve the steps for a recurring job so that the assistant does not have to work them out again. By November, the two products overlapped considerably: both could run scheduled work, use external services and keep information for later sessions. The stable releases inspected were OpenClaw v2026.11.3 and Hermes v0.23.1, both published on GitHub on 10 November [5], [8].

Either can become a personal assistant or an automation system. Choosing one requires looking at the particular routine, the tools it needs and the amount of maintenance its owner is prepared to do.

Email triage and daily briefings

Radek Sienkiewicz's account of fifty days with OpenClaw describes email checks and replies prepared for his review. The agent flags time-sensitive items, and he sends the messages himself. Its background checks have surfaced payment failures, renewal deadlines and upcoming meetings, while a separate morning digest collects material relevant to his interests [9].

The result can be checked against the original messages. An alert should identify the sender, the deadline and the reason it matters. A draft should be visibly waiting for approval. A morning briefing should point to the calendar event or email that supports each item. These are recommended acceptance rules for adapting the workflow, rather than features that an installation automatically enforces.

Hamza Mostafa describes a similar OpenClaw arrangement on a Mac mini: scheduled briefings, email sorting and preparation for meetings, with Telegram as the place he issues instructions. When he asks for a reply, the agent reads the thread and shows him the draft. He also describes something the successful screenshots omit: scheduled jobs ran for two days, incurred model charges and produced output that never reached his Telegram chat. He reports about $15 in wasted API calls before finding the delivery-setting problem [10].

A delivery test would have exposed the problem earlier: a known email should appear in the next briefing, and a prepared draft should arrive in the agreed place. The owner also needs a notice when an account cannot be checked.

For an initial setup, connect one inbox and one calendar, then ask for a short digest at a fixed time. Restrict it to appointments, messages requiring a reply and deadlines. Keep replies as drafts until repeated use shows that the assistant understands the relevant people and commitments. Sienkiewicz recommends email triage, a daily briefing and a place to collect bookmarks as starting workflows; he also reports that complex browser tasks still require nudging [9].

A briefing can replace repeated visits to the inbox and calendar. Someone with five emails a day may gain very little; someone who spends a morning reading a crowded inbox and checking several calendars has more work to delegate. Compare the time spent checking the briefing with the previous routine. Neither account supplies a general productivity measurement.

Sienkiewicz also describes collecting links in a dedicated Discord channel. He drops a bookmark there, and the assistant enriches it and stores it with his notes. This is useful for someone who gathers material throughout a week but later cannot remember why a link mattered. The result should preserve the address, a short description and the reason it was saved. In his setup, separate channels keep research, analytics and everyday tasks from becoming one sprawling conversation [9].

That separation has an ordinary practical benefit: the person knows where to look. A channel for client-meeting preparation can contain the relevant notes and pending questions, while a household channel holds appointments and shopping. It also gives a clearer instruction than “remember everything.” A useful saved note can be corrected or moved; an assistant's vague assurance that it remembers is difficult to inspect.

Utility bills and household documents

TekLoon runs OpenClaw locally on a Mac and communicates through WhatsApp. On the fifth of the month it checks for missing utility bills, searches the relevant Gmail senders and downloads the attachments into the appropriate month in Dropbox. On the twentieth, a separate routine finds the bills for the required accounting period, extracts the amounts and prepares the payout request [1].

The system includes an important business rule: the request concerns a month three months earlier. Without that instruction, an assistant could produce a perfectly formatted message for the wrong period. A small calculation program handles the arithmetic, while OpenClaw connects the steps and delivers the result. TekLoon forwards the prepared message himself. His account supports bill collection and message preparation; it does not describe autonomous payment handling.

TekLoon spent a few hours teaching a routine that previously took about fifteen minutes each month. Setup therefore outweighs the initial time saving. The bills now reach the right folder without another reminder, and the payment request is prepared on schedule. His report does not measure the time he spends checking or maintaining it [1].

Arthur Soares offers another OpenClaw document workflow from his household in Berlin. His family keeps contracts, receipts, invoices and insurance policies in Paperless-ngx, a document-management application. The assistant helps categorize incoming documents and answer questions about them. During a commute, he asked by voice for information about legal insurance. The agent located the policies, summarized them, noticed that they still contained an old address and drafted a correction email for his approval [11].

Soares can ask about a policy while travelling because the assistant can consult his document collection. His same account describes memory mistakes, interrupted context and scheduled jobs that misfire; the policy file lets him check a reply against the written terms [11].

Keep the original bill or policy beside the assistant's answer, with its date and the relevant page. A missing attachment should be flagged. If a number looks wrong, the owner can then check it without repeating the document search. These are suggested checks for adapting the reported routines.

Invoice approvals and staff reminders

Yash Gadodia's May account describes OpenClaw doing work inside a company. His setup included customer-facing WhatsApp groups, a Xero invoice-approval flow, morning updates, unpaid-invoice summaries and staff reminders. He reports seven agents and approximately thirty scheduled jobs on an office Mac mini [12].

Gadodia changed the unpaid-invoice routine after it sometimes altered an amount or omitted a customer. Initially, an AI instruction asked the assistant to inspect the invoices, write reminders and send the results. He replaced that part with a fixed program that reads Xero and formats the report consistently. The assistant continued to handle intake and communication, while the accounting records supplied the numbers [12].

Gadodia reports failures that were hard to notice, confusion over which bot sent a message, and a messaging connection that needed automatic recovery. He made failed scheduled jobs send him direct messages so that he could repair a broken routine before a customer asked about it.

For an invoice workflow, the sensible division is explicit: the agent can collect a request, retrieve the customer record and prepare an invoice for approval; the accounting application determines its status and amount. Sending a reminder or changing a record should follow the business's existing approval rule. This is an adaptation of the reported approach, not a claim that every OpenClaw installation includes a finished Xero workflow.

The same principle applies to internal reminders. Birthdays, leave requests and missing daily updates can be useful scheduled messages because the rule is limited and the outcome is observable. Letting a model decide which customer can be ignored, or how much a customer owes, creates a different responsibility. Gadodia's experience demonstrates why those decisions should remain tied to the business's records [12].

Gym bookings and daily menus

Macedo's gym opens reservations twenty-four hours before each class, and popular sessions fill quickly. His Hermes setup runs on a home Mac mini and messages him through Telegram. At eight in the morning, a program fetches the next day's timetable, removes class types he dislikes, checks calendar conflicts and excludes existing bookings. Hermes receives the remaining choices and asks which he wants [2].

He can answer with a short selection, such as the first and fourth options. The system then schedules the actual booking for the minute each reservation window opens. A confirmed place creates a calendar event; a full class leads to a waiting-list request instead. He describes a morning selection followed by an evening booking without opening the gym app himself.

The booking routine removes the need to watch the reservation window. Macedo supplies his preferences, the club supplies availability, and a calendar event confirms that he has a place.

The first version let the model perform all the checking. Macedo reports roughly 700,000 tokens per morning to produce twelve class options, eventually hitting a provider spending limit. The revised version performs the checks in a program and asks the model to present the result and handle the reply. His reported 6,000-token morning exchange shows a large reduction within that particular setup [2]. It does not establish that Hermes generally uses fewer tokens than OpenClaw.

There are still limits. Joining the waiting list does not guarantee a place, and the system did not yet handle every notification when a place became available. Macedo also found an old waiting-list entry that had not been cleared. A booking assistant needs to distinguish “requested,” “waitlisted” and “confirmed”; reporting all three as success would hide the most important part of the job.

His restaurant-menu routine is another revealing example. It reads a daily menu posted as an image, matches dishes to his saved ratings and tells him only about food he is likely to want. The first attempts sometimes served yesterday's menu because the restaurant had not replaced the image. He added a check of the date and repeated checks for an updated image. He reports about 40,000 tokens a day for this image-reading job, and occasional duplicate requests to rate a dish whose name was written differently [2].

A simple-looking menu check can consume more model usage than a booking routine with carefully filtered data. Choosing an agent is only one part of choosing a working solution. The format of the source, the frequency of checks and the amount of unnecessary reading can matter just as much.

Household expense tracking

Maish Saidel-Keesing built a Hermes budget bot that he and his wife can use through WhatsApp. Credit-card purchase messages supply the transaction data. The system records expenses, tracks categories and answers questions in Hebrew or English. Its central rule is that ordinary software calculates the balances; the AI interprets a question or suggests the category for an unfamiliar merchant [13].

A new merchant's charge does not count against a guessed category until the user confirms it. Once a rule is established, later transactions can use it. Some merchants remain “ask me” cases because they sell several kinds of goods, and a shop name alone does not identify what was purchased.

The account also contains failures worth understanding before copying the idea. At first the bot ignored messages because the allowed-user setting was in the wrong place. Later it replied conversationally without using the budget tools. A one-word response to a notification confused it because it could not see the notification in the same conversation. These are usability problems: the person thinks they answered the assistant, while the assistant lacks the information needed to interpret the answer [13].

Saidel-Keesing addressed them by narrowing the bot's role and making a bare reply refer to the newest pending charge. He gave the family bot only its budget functions, keeping it separate from his more powerful personal assistant. He reports that a reply on a free-tier model took eighty-five seconds, even though recording incoming charges did not depend on that slow conversation.

The distinction between building and using the system is also important. Claude Code helped build the bot; Hermes ran it. His report of getting the core implemented quickly is not evidence that Hermes independently developed the entire solution. Refunds and instalments were still untested against real messages, and he planned to compare it with his paper records for two weeks before retiring them [13].

For readers considering this use case, a budget summary is a more defensible starting point than giving an agent payment authority. Import a limited set of expenses, confirm ambiguous categories and reconcile the totals with a statement. The convenience lies in asking “how much is left for groceries?” without opening a spreadsheet, while keeping the calculation and the records inspectable.

Lead research and client meetings

Halfbit Studio, a three-person software company in Poland, describes Hermes helping with competitor analysis, lead research, sales-message revision and finding relevant events. Marcin Golonka reports research on twenty-four leads, with company profiles and first-contact recommendations recorded in YouTrack, the team's work-management system. He also describes saving the research procedures so that they can be repeated without rebuilding the company's background each time [14].

The concrete output is a prepared brief and a place to act on it. A sales person can inspect the company's activity, verify the proposed contact, check the evidence for a business problem and decide whether to reach out. Halfbit says it retains responsibility for client communication and decisions; the agent supplies preparation.

Halfbit reports faster preparation of competitor reports and acknowledges incomplete research and generic recommendations. It does not provide a measured increase in sales or an independently checked financial saving [14].

Jim Stewart's agency account supplies a more specialized Hermes example. He describes agents with client context helping prepare meetings and investigate Shopify search issues. One workflow diagnosed a structured-data problem, proposed a fix, backed up the site and applied the change. He subsequently observed Google Shopping crawler activity, while explicitly describing the connection as correlation requiring further confirmation [15].

Stewart's process includes a diagnosis, a proposed change, a backup and a check of the site afterward. He was the only person in the office using Hermes at that point, and building and tuning the agents made him busier. His report describes his own use and a site correction, without establishing a sales improvement.

Both cases show why “run my company” is an unhelpfully broad starting request. A lead brief with source links, a meeting summary assembled from existing notes, or a proposed site correction is a result a person can judge. Invented prospect details, unsupported sales claims and changes to a live customer site require much more scrutiny than the polished wording of the response suggests.

For a small company copying Halfbit's approach, the initial deliverable could be a single-page brief for one prospect: what the company does, the public evidence for a relevant problem, the proposed contact, and questions that remain unanswered. Save the brief with the lead, then draft a message using only facts the owner has checked. Repeating this for three known companies is more informative than asking for hundreds of contacts on the first run. A large output can multiply incorrect details just as efficiently as correct ones.

Meeting preparation is similarly concrete. Give the agent access to the previous meeting notes, the agreed actions and the current project record. Ask for the unresolved decisions and the evidence needed for each. The valuable output is a short agenda that helps the meeting start with the right information. A summary that merely sounds polished, or invents progress because it cannot find an update, adds little.

For these suggested research tasks, ask the assistant to flag missing evidence and separate a prospect's statements from inferred needs. Check the brief against the original material before using it in a meeting or an outreach message.

Marketing and social media

OpenClaw with Buffer across several brands

Victoria Lo describes an OpenClaw publishing setup for ragTech, her personal writing and Women Devs Singapore. She connected the assistant to Buffer, a social-media scheduling service. Her ragTech channels include a LinkedIn Page, an Instagram Business account and TikTok; her writing has a LinkedIn profile and a separate Page; Women Devs Singapore has its own LinkedIn Page [16].

She asks the assistant to prepare an event announcement, promote a podcast episode or share an article. It drafts the caption using the selected brand's writing rules and schedules it to the appropriate account. ragTech's captions have a conversational style, while her personal LinkedIn posts use a more direct professional tone. A request specifying Thursday at eight in the morning, Singapore time, supplies the publication time. The setup saves her from opening separate accounts and uploading each post manually [16].

Her recurring “Techybara Explains” series uses pre-drawn mascot artwork and a font made from her handwriting. The agent prepares the text, and a template combines it with the existing artwork to make carousel slides. The finished slides are scheduled across ragTech's LinkedIn, Instagram and TikTok channels. A related article becomes a newsletter draft, which Lo reviews before sending. She still chooses the topic, edits the script and approves the schedule [16].

Lo reports that the earlier manual carousel routine took one to two hours per edition. She also reports more than three thousand views for the recurring series before describing its automation. Those views show an audience for the series; they are not evidence that OpenClaw created three thousand new followers or increased reach after installation. She describes the revised routine as reducing the burden on her schedule, without publishing a measured before-and-after production time [16].

Hermes with a publishing service

Jim Stewart says Hermes prepares social content from the agency's work and can make accompanying short videos. He also says the writing sometimes sounds too artificial and the videos have layout problems [15]. An agent can prepare material for several channels, but someone still has to supply the work worth discussing and edit the result.

Blotato founder Sabrina Ramonov documents a Hermes demonstration by Zach Babiarz. He gave Hermes a link to his own YouTube video and requested an Instagram post and an X post. The assistant read the transcript, prepared different captions and an image, and kept them as drafts. After he approved publication, it sent the posts through Blotato and checked the published addresses [17]. This is a vendor-documented demonstration; it is not an independent campaign result.

Postiz documents another connection for Hermes and OpenClaw, with account discovery, drafts, scheduled posts and performance reporting. Its Hermes guide includes Instagram, LinkedIn and Facebook, and describes connecting the social accounts inside Postiz before asking the agent to use them [18]. These services perform the platform-specific publishing. Installing either agent alone does not connect every social account or provide permission to post.

For a person whose main need is a weekly posting calendar, the scheduler itself may be sufficient. Adding an agent is more relevant when the material also has to be collected from meeting notes, articles, product updates or videos, rewritten for different audiences and returned for approval.

Instagram, LinkedIn and Facebook accounts

Account type affects what a publishing service can do. Buffer supports multiple accounts from the same network, so several client Pages or brand accounts can appear in one workspace. It also lists restrictions: Facebook personal profiles cannot be connected, and some channels use reminders that require a person to finish publication [19].

A multi-account request should identify the brand, destination and time. “Publish this everywhere” leaves too much unresolved when the workspace contains a founder's profile, a company Page and client accounts. The approval message should show the account names as well as the captions.

For example, a company promoting a training event could ask for an Instagram carousel showing what participants will learn, a LinkedIn announcement explaining who the event is for, and a Facebook Page post with the date and registration link. This is a suggested campaign, not an additional reported deployment. The event details should come from one approved brief, while each version is edited for its destination.

The work can be divided into five steps:

- Collect the approved topic, source material, event or product details, and permitted images. - Prepare a caption and media selection for each named account. - Return the account names, previews, links and publication times for review. - Schedule the approved versions and check whether each is scheduled, published or failed. - Record the live links and errors so that a failed upload can be retried without duplicating a successful post.

Lo's account demonstrates separate writing rules for several brands, while the Hermes demonstration shows drafts followed by approval and publication checks [16], [17]. The sequence above adapts those practices to a campaign on Instagram, LinkedIn and Facebook.

Followers, engagement and customer enquiries

An agent can help maintain a posting schedule, prepare replies and collect performance figures. It cannot establish demand for a product merely by publishing more often. None of the marketing accounts inspected for this review provides a controlled result showing that OpenClaw or Hermes, by itself, produced a particular increase in followers.

To work on audience growth, give the assistant a defined audience and a set of approved posts, comments and customer questions. It can group repeated questions, propose a series answering them, draft variations and prepare the next week's calendar for review. For the training-event example, that could mean explaining a common problem, showing an exercise from the course and answering an attendee's question with permission. Compare the responses to those posts before repeating the topic or increasing the posting frequency. This is a suggested editorial process; its effect on followers has to be measured on the account.

Buffer's Insights documentation describes tracking follower growth, engagement and impressions across supported channels [21]. Its separate explanation of data differences notes that historical coverage varies by network, figures are collected in batches and some results can take up to twenty-four hours to appear [22]. A report should identify the date range and any missing data before comparing accounts.

For a weekly review, ask for:

- Audience growth: starting and ending follower counts for each account, with losses included. - Content response: reach or impressions, saves, shares and substantive comments, using the metric names supplied by each platform. - Traffic: visits to the linked article, product or registration page, using identifiable campaign links. - Business results: enquiries, registrations or purchases that can be connected to the campaign. - Work required: time spent editing, approving and repairing posts.

These are suggested measurements, not results observed by EFFOMA. Views, followers and enquiries answer different questions. A widely viewed post may attract few followers; an established small account may generate enquiries without a large increase in audience. The agent's report should retain those distinctions.

Comment and message management also depends on the connected service and platform permissions. A publishing connection should not be assumed to include every inbox. Start with collecting unanswered questions and drafting replies. Complaints, refund requests, personal information and promises about delivery or pricing need the person responsible for those decisions. Repeated generic comments, unsolicited messages and invented customer stories are poor material for building an audience, however easy they are to generate.

Phone requests and review

An assistant connected to a computer can be useful while its owner is elsewhere. Sienkiewicz describes asking OpenClaw from his phone to make a small code change and prepare it for later review. He explicitly says this is for quick fixes and simple ideas; his main programming work still happens with dedicated coding tools [9].

For a non-programmer, the relevant capability is delegating a bounded job without opening the laptop immediately. A voice note can ask for the invoice files needed tomorrow, a meeting brief, or a proposed website correction. The agent needs access to the actual records, and the result should arrive as a file, a draft or a visible change for review. “Done” in the chat is a status claim until the owner can inspect the output.

This distinction becomes important when work crosses several applications. An agent may write a file successfully but fail to upload it, prepare a message without delivering it, or update a local copy while leaving the shared version unchanged. Mostafa's missed briefings and Gadodia's customer-facing maintenance problems show why the final destination has to be part of the definition of completion [10], [12].

The reply should identify the output: a draft in the agreed folder, a booking confirmation or a link to the updated document. It should also identify any decision still awaiting approval.

Workflow comparison

The following table separates reported user practice from a recommendation about fit. A documented example proves that an owner assembled a particular workflow; it does not mean it appears fully configured after installation. Both agents may perform many of these jobs when connected to the appropriate services.

The absence of a matched example in this review is a limit of the evidence, not proof that a feature is unavailable. Neither product has a convincing general claim to be the best at every task.

November releases

OpenClaw's Multiplayer announcement describes several people contributing to one shared conversation, seeing the same context and steering ongoing work [24]. A household member or teammate can see the earlier instructions when taking over a task. Conflicting instructions remain possible, so a shared chat needs an owner for decisions involving money, customers or appointments.

Hermes's November update brings its desktop, browser, editor and messaging sessions under a common service, helping users continue work between interfaces. It also adds approval and running-total controls for paid external tools that declare their prices [23]. Those controls are useful when research calls a charging service, although they are not a universal ceiling on every model and hosting bill.

OpenAI's 12 November report describes more than five hundred persistent agents internally and the organizational controls used around them. It says OpenClaw Enterprise remains a pilot, with a first production version planned for the end of the year [25]. This provides evidence of substantial internal use, but not a published count of hours saved or a reason to expect a home installation to behave like OpenAI's managed deployment.

These improvements concern operating an assistant over time. Neither release establishes a general performance advantage in ordinary users' work, and connected accounts still need to be configured for each routine [5], [8], [23]–[25].

Running costs and recurring problems

There is no defensible single monthly price for OpenClaw or Hermes. The software is openly licensed, while a user's expenditure depends on the model, paid searches and other services, the machine running it, the frequency of background tasks and the amount of repair. Mostafa's March OpenClaw account places his combined recurring setup around $120–195 a month, including model use, search, access to X data and a dedicated phone number. The Mac mini was a separate purchase [10]. That is his historical budget, not a November quote.

Macedo's Hermes experience demonstrates a different issue: poorly designed routine work can consume far more model input than expected, even before it produces anything valuable. His 700,000-to-6,000-token change came from changing the workflow [2].

A useful household assistant should also be quiet when nothing matters. Macedo's routines avoid messages when there is no relevant change, and Gadodia emphasizes direct notification on failures [2], [12]. Constant “all clear” messages can turn a supposed time saver into another notification stream to manage.

The difficult tasks remain the ones with many uncertain steps: unfamiliar websites, changing logins, ambiguous documents, conflicting instructions and decisions where a mistake is expensive. Sienkiewicz reports unreliable browser sessions and work that stops until he asks for progress; Soares reports confident memories of things that never happened [9], [11]. Saved knowledge can reduce repeated explanation, but consequential facts still need a document, record or confirmation behind them.

Installation and first-week checks

Start with one briefing, document routine, booking or approved batch of social posts. Each has an identifiable input and a result that can be checked before more accounts and scheduled jobs are added.

For a briefing, the initial instruction can be specific: “At eight each weekday, send today's appointments and messages needing a reply. Include links to the originals. Prepare replies as drafts and tell me if an account could not be checked.” For bills: “Collect attachments from these senders, file them by billing month, prepare the usual message, and flag any missing or unreadable amount.” These are suggested instructions based on the reported workflows, not claims that prompts alone create or secure the required connections.

Before commissioning a setup, specify the connections and the output in everyday language. The examples below describe the work a configured system should perform; they are not additional observed deployments.

There is little reason to buy dedicated hardware before this small test works. The reported cases use machines their owners can keep running, including Mac minis, but those anecdotes do not establish that a particular brand of computer is necessary [1], [2], [10], [12], [13]. A machine that is switched off cannot execute its local morning job; a hosted installation needs somebody to preserve its records and maintain its connections. Where it runs is an operational choice, not a measure of how intelligent the assistant is.

Someone still needs to install the software, connect the accounts, set permitted actions and verify the schedule. Tessa Kriesel's Hermes deployment account describes a bot that lost its settings on redeployment and another startup problem that left it unable to answer. After a restart, check that the previous settings remain and that a test message receives a reply [26].

For the first week, compare the output with the task as done manually. Record delivery, errors, review time and the provider's charges. If repair takes longer than the previous routine, reduce the task's scope. Add another routine after the first one delivers consistently.

For social publishing, the unresolved work is often between the approved draft and the final post: an expired account connection, rejected media or a schedule interpreted incorrectly. The owner needs a record of what went live, what failed and what still awaits approval. Neither a longer content calendar nor a confident completion message substitutes for that record.

References

[1] TekLoon, “Automating My Rental Property Bills with a Personal AI Assistant,” Feb. 22, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://tekloon.net/blog/tech/automating-rental-bills-with-ai/

[2] H. Macedo, “Three things my agent does for me every day,” Sep. 9, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://ickas.dev/writing/three-things-my-agent-does-every-day

[3] P. Steinberger, “Introducing OpenClaw,” OpenClaw Blog, Jan. 29, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://openclaw.ai/blog/introducing-openclaw

[4] P. Steinberger, “OpenClaw, OpenAI and the future,” Feb. 15, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://steipete.me/posts/2026/openclaw

[5] OpenClaw contributors, “OpenClaw,” README and release notes, tag v2026.11.3, Nov. 10, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://github.com/openclaw/openclaw/tree/v2026.11.3

[6] D. Morin, “Introducing the OpenClaw Foundation,” OpenClaw Blog, Jul. 8, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://openclaw.ai/blog/introducing-openclaw-foundation

[7] Nous Research, “Releases,” Hermes Agent launch entry, Feb. 25, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://nousresearch.com/releases

[8] Nous Research, “Hermes Agent v0.23.1,” release notes, Nov. 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.1

[9] R. Sienkiewicz, “50 days with OpenClaw: The hype, the reality and what actually broke,” VelvetShark, Feb. 26, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://velvetshark.com/50-days-with-openclaw

[10] H. Mostafa, “I Run a Personal AI Agent 24/7 on a Mac Mini. Here's How It Actually Works,” Mar. 7, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://hamzamostafa.com/blog/openclaw-setup

[11] A. Soares, “… with all its claws,” Mar. 14, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://arthur.earth/posts/with-all-its-claws/

[12] Y. Gadodia, “OpenClaw, six months later: what production actually demanded,” May 11, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://yashgadodia.com/openclaw-six-months-in-production/

[13] M. Saidel-Keesing, “I Replaced My Paper Budget With a WhatsApp Bot. Here's What Broke,” Technodrone, Oct. 5, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://blog.technodrone.cloud/2026/09/family-budget-whatsapp-bot.html

[14] M. Golonka, “How We Run Our Company with Hermes Agent,” Halfbit Studio, Jun. 22, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://halfbitstudio.com/en/how-we-run-our-company-with-hermes-agent/

[15] J. Stewart, “AI Agents Are Changing How We Run Our Agency,” StewArt Media, May 11, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://stewartmedia.com.au/ai-agents-are-changing-how-we-run-our-agency/

[16] V. Lo, “Building My Second Brain on OpenClaw (Part 9),” Aug. 17, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://lo-victoria.com/building-my-second-brain-on-openclaw-part-9

[17] S. Ramonov, “How to Make Hermes Agent Post to Social Media (All Platforms),” Blotato, Jul. 21, 2026, updated Oct. 26, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://www.blotato.com/blog/hermes-agent-social-media

[18] Postiz, “Schedule Social Media Posts with Hermes Agent,” updated Sep. 25, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://postiz.com/hermes-agent

[19] Buffer, “Supported channels,” Buffer Help Center. Accessed: Nov. 13, 2026. [Online]. Available: https://support.buffer.com/en-us/articles/supported-channels-LM3P7Y4zsp

[20] Buffer, “Using Facebook with Buffer,” Buffer Help Center. Accessed: Nov. 13, 2026. [Online]. Available: https://support.buffer.com/en-us/articles/using-facebook-with-buffer-jY1AM9fz2G

[21] Buffer, “Getting started with Buffer's analytics features,” Buffer Help Center. Accessed: Nov. 13, 2026. [Online]. Available: https://support.buffer.com/en-us/articles/getting-started-with-buffers-analytics-features-UVAe3hwLD8

[22] Buffer, “Why your data in Buffer Insights might look different from native analytics,” Buffer Help Center. Accessed: Nov. 13, 2026. [Online]. Available: https://support.buffer.com/en-us/articles/why-your-data-in-buffer-insights-might-look-different-from-native-analytics-CryTKTdV0u

[23] Nous Research, “Hermes Agent v0.23.0,” release notes, Nov. 5, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://github.com/NousResearch/hermes-agent/releases/tag/v0.23.0

[24] J. Pishgar, “Shattering Silos with OpenClaw Multiplayer,” OpenClaw Blog, Nov. 5, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://openclaw.ai/blog/shattering-silos-with-openclaw-multiplayer

[25] OpenAI, “Securing persistent agents like OpenClaw at OpenAI,” OpenAI Developers Blog, Nov. 12, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://developers.openai.com/blog/securing-persistent-agents-like-openclaw-at-openai

[26] T. Kriesel, “How I Deployed Hermes Agent on Railway with Telegram (And Every Gotcha I Hit Along the Way),” May 12, 2026, updated May 13, 2026. Accessed: Nov. 13, 2026. [Online]. Available: https://tessakriesel.com/how-i-deployed-hermes-agent-on-railway-with-telegram-and-every-gotcha-i-hit-along-the-way/