273 AI Agents in One Day: How Our Agent-Run Company Upgraded Itself
This is an adapted English version. Original (Russian): automata.sale/blog/ai-automation/273-agenta-v-odin-den/
273 AI agents in one day is not a clickbait headline, it is our internal CRM morning report. On Monday I found a link to an open catalog of agent prompts. By noon, three subagents had taken it apart file by file. By evening, nine company roles were operating under new rules. Not a single human read any of those 273 files in full. Here is how a company run by agents upgrades itself, and what you can replicate in your own business tomorrow.
What we found: a library with 169 thousand stars
agency-agents is one of the most hyped repositories of the year: 273 agent files in 22 “divisions”, from finance and sales to game development and spatial computing, almost 169 thousand GitHub stars, MIT license. Inside each file sits a ready-made role: an identity, a mission, hard rules, document templates, a workflow and success metrics.
First, an honest observation: this is a prompt library, not a system. There is no memory between runs, no scheduler, no gates, no access to the live data of your business. It is a construction set you can build an operations layer from, but the catalog itself runs nothing.
Second: 90% of the content is noise for us. Divisions for Douyin, WeChat and Unity, “whimsy injectors” and carousel generators for platforms blocked in Russia. But the remaining 10% is gold: methodologies nobody invents from scratch in a day, because hundreds of practitioners polished them first.
And this is where it gets interesting. To mine gold out of 273 files you do not need a week and a pot of coffee. You need to delegate the reading to agents.
Context: a company that could run this experiment
For those meeting us for the first time. Automata CRM is a system where an Obsidian vault plays the role of the single database: clients, invoices, tasks, tickets and decisions live in plain markdown files. Agents work on top of that:
- Five scheduled roles: a commercial director touches clients every morning, a finance director counts numbers every Friday, a CEO sets the weekly bar on Mondays, a marketer runs the content pipeline biweekly, an on-call SRE walks the systems every evening.
- An hourly inbound cycle: email, Telegram and MAX are processed every hour; a client gets an answer in minutes, not days.
- A company charter with hard gates: an agent physically cannot send an invoice below the price floor, promise a deadline without owner confirmation, or publish a draft with a discount above 10%.
- A board of directors: a dashboard where every role reports, and questions to the owner turn into buttons. A decision takes one click instead of a thread of messages.
The key principle: a role is not an ever-living chatbot with a “personality”, it is a checklist (a skill) plus a schedule slot plus charter rules. That is why upgrading the company never looks like “let us plug in twenty more agents”. It looks like surgical replacement of methodologies inside roles that already work.
The extraction mechanics: three subagents, one pass
We cloned the repository locally: 273 files, conversion scripts excluded. Then three subagents launched in parallel, each with its own zone:
- Marketing, sales, paid media: 53 files.
- Finance, project management, support, product, strategy: 40 files.
- Testing, engineering, specialized roles: about 70 more.
Each had the same task: read everything and return a verdict per file in the context of our system: “steal / skip / reference”. Not a summary, a decision: what exactly to take and where exactly to put it.
By midday I had distillates on my desk: verdict tables covering 160+ files, deep breakdowns of the top five in each zone, and a ready implementation plan. A human would have read for a week. The subagents finished in a morning and spent my attention only on the final calls.
Then came the surgery. The edits did not go “into the AI in general”, they went into concrete role skills. And since our automation prompts reference the skills, the updates picked themselves up: from the next run, the rituals worked under the new rules. Nothing was redeployed, nothing restarted.
Money: a dunning ladder and one-button decisions
The most valuable find was not even a methodology but a behavioral design principle: every question to the owner must arrive with a ready default answer. Not “what do we do about this debtor?” but “the letter to the debtor is drafted, send or edit?”. Consent costs one click, and default consent breaks procrastination, the disease that nine out of ten small businesses suffer from, especially in money matters.
Around that principle our commercial director got a full receivables ladder:
| Day | Action | Format |
|---|---|---|
| D-3 | ”Invoice due in 3 days” | Sent immediately |
| D+1 | Reminder with PDF invoice | Sent immediately |
| D+7 | Serious reminder with the acceptance act | Sent immediately |
| D+14 | Formal demand + “do we pause work?” | Draft for confirmation |
| D+30 | Pre-legal claim | Via the lawyer |
The finance director now computes things a classic small-business report never had:
- RFM health score for every client with Champions, Loyal, At Risk segments and the “Cannot Lose Them” category: a client who went silent but used to pay a lot is the profile of a retainer about to churn, and it goes into the daily touch plan immediately, not a month later when it is too late.
- A 90-day cash flow with a minimum balance threshold. Threshold breached: the prescribed action is “accelerate receivables or delay payables”, not “think about it”. The difference between a report and management.
- Churn revenue impact: how much recurring revenue the departed took with them this month. A metric even larger companies usually lack.
- The report ends with actions, not numbers: “immediate / 30 days / strategic”, each with an owner and a date. The CFO KPIs: less than 5% of receivables past 90 days, cash forecast accuracy within 10%.
None of this requires programming. These are rules you can put into any CRM tomorrow, if you have someone who never forgets to execute them. We do.
Quality: the end of fantasy reports
The second catch came from the testing division, and it hits the main pain of everyone working with AI: fantasy reports. “Checked, everything works” is the most dangerous phrase in the agent world. An AI states a real check and a hallucinated one with equal confidence.
Our QA skill got an evidence protocol:
- Every “it works” claim requires proof: a screenshot, a curl output, a log, a green run. No proof, status NOT TESTED.
- Four statuses instead of two: PASS (proven), FAIL (proven broken), NOT TESTED (a hole in the testing, not “all good”) and NOT DETERMINED.
- Automatic fail for “no issues found” without a documented test scope. A score without a rubric is a fail. “Production ready” without a list of what was checked is a fail.
- The default verdict for any fresh work is NEEDS WORK: first implementations take two or three revision cycles, and that is normal, not a disaster.
- Perfect scores from a previous agent are a red flag, not a green light: reproduce two or three points at random.
One more quote worth framing, from their reviewer prompt: “a prompt is a spec. If the model did not do what you wanted, the spec was ambiguous, not the model’s fault.” That applies to any instructions you give human employees, not only to AI.
The content pipeline: citation measurement instead of guessing
We have done AEO for a while, optimizing content for how ChatGPT, Perplexity and Yandex Neuro answer. But the catalog had the missing link: a closed citability measurement loop.
Once a month: collect 20-40 live prompts that real people ask AI (“recommend a Bitrix integrator”, “bitrix vs amo”), run them through four platforms and record a table of “they cite us / a competitor / nobody”. Lost prompts get a separate analysis: why the competitor won, what they have (a comparison page, FAQ schema, freshness). Then a fix pack and a re-check with the same prompt set in 14 days. Benchmark before fixes, always, or the effect is invisible.
The content plan itself now builds from live query formats: “best X for Y” becomes a comparison, “X vs Y” a dedicated comparison page, “how to choose X” a guide with a selection framework, “how much does X cost” a breakdown with numbers.
Plus three hygiene rules almost nobody follows: a cannibalization blocker before writing a new article (two articles targeting one intent means deciding first who owns the query), a token budget per page (a landing page stays under 8k tokens of content, an AI crawler with a limited context simply will not finish longer walls of text) and pipeline memory: a learnings file with hook formulas and posting times that the agent reads before every generation, to repeat winners instead of stepping on the same rake.
Sales: six questions you cannot skip
From the sales division we took four things, each the size of a paragraph but working like a process:
Six mandatory qualification questions. What is broken (in the client’s words)? Why is it broken (the root)? What does it cost (money, time, risk)? Who else cares? Why now? What happens if they do nothing? A lead without answers to these six is underqualified and does not go into the forecast. That is it.
The “expensive” objection is almost never about money. The AECR scheme: acknowledge, empathize, clarify (“when you say ‘not now’, is that budget, resources or something else?”), reframe. In their field data, nearly half of objections are about misunderstood value and a third about timing; only the remainder are genuinely about budget.
The stalled-stage rule. A deal sitting on one stage beyond 1.5x the median stage duration gets one touch with a new angle or moves to a soft nurture track. Nothing hangs silently.
The offer auditor, a value equation. Value = (dream outcome x perceived likelihood) / (time delay x effort). Before cutting the price, squeeze the other three levers: prove the likelihood, shorten time to first result, remove effort from the client. That is almost a consulting engagement in one formula.
Why we did NOT install 273 agents
This needs to be said plainly, because there is a lot of noise around such catalogs. We deliberately did not plug in a single agent from the repository as an agent. Reasons:
- A personality without memory is not an employee. Those 273 prompts do not know our clients, our prices or our discipline. Our strength is the bundle: skill + schedule + charter gates + Obsidian as the single source of truth.
- Ever-living agents are expensive and chaotic. Every role that “lives” idle burns attention and budget. Our roles wake up in their slot, do the work, write the journal and go back to sleep.
- Format matters less than content. Of their file structure we took three sections: hard rules (numbered Never/Always), a ready-made deliverable template and verifiable metrics. The “identity” and the “vibe” are cosmetics we declined.
The summary of this philosophy: we do not run 273 agents. We run five roles that get smarter every week, because every find gets stitched not into a chat but into the operations layer.
What this case shows your business
Three takeaways worth the whole article:
One. The competitive edge of 2026 is not access to AI (everyone has it), it is the speed of absorbing methodologies. The world produces best practices faster than any human can read them. A company with reading and adoption on a conveyor belt upgrades weekly; a company without one never does.
Two. Agents do not replace the owner, they prepare decisions for him. Every question our system asks the business owner arrives with a ready default: a drafted letter, a button, a date. One letter instead of an hour of deliberation. That is running a company in “approval mode” instead of “decision production mode”.
Three. Honest systems beat pompous ones. We rejected dozens of beautiful prompts tagged “+300% leads”: numbers out of thin air. Methodologies with verifiable rules (proof for every claim, KPIs with thresholds, a re-check in 14 days) went into production the same day. Apply the same filter to any “miracle tool”: if a metric cannot be verified, it does not exist.
What is next
Seven upgrades already sit in the review queue: a monthly financial review, quarterly business reviews with clients, charter escalation updates, win themes for proposals, contract clauses, a code drift registry and WebMCP readiness for the agent era. Some get approved this week, some go to the backlog, and that is fine: in an agent-run company the backlog does not rot, because every Monday the CEO role asks for a report on it.
If you want to see how it all looks live, the Automata CRM dashboard is open for a demo, and we regularly dissect the internal kitchen in this blog. The question “how are you different from an agency where humans do everything by hand” now has a concrete answer: our operations layer can upgrade itself, and you have just read the report of one such upgrade.
📞 +7 (906) 311-77-69 · ✉ hello@automata.sale · 💬 Telegram: @automatasale · 🌐 automata.sale
Evgeny Uryadov, sole proprietor (INN 645112058391, OGRNIP 312645301900058) Working across Russia, Belarus and Kazakhstan
Frequently asked questions
What is agency-agents and why did it matter?
It is an open GitHub catalog with almost 169 thousand stars: 273 agent files across 22 divisions, from marketing and sales to finance and testing. Each file is a ready-made role with rules, deliverable templates and success metrics. We ran it through three subagents and absorbed the best parts into our own operations.
How is your system different from just a set of prompts?
A prompt is a personality without memory. In our setup a role is a skill with hard rules plus a scheduled slot plus gates from the company charter: an agent cannot send an invoice below the price floor or promise a client a deadline without owner confirmation. Data lives in an Obsidian vault, and every role reads the live state of the business.
How many agents does a small business actually need?
Not 273 and not 30. Our working roster is five roles (commercial director, finance director, CEO, marketer, on-call SRE) plus a lawyer on demand and an hourly inbound cycle. Each role wakes up in its slot, does the work, writes a journal entry and goes back to sleep.
What did this operations upgrade cost?
Zero rubles beyond the already running setup: the audit and the edits were done by the agents themselves in one working session. The expensive part is not extracting ideas but maintaining the system around them: the schedule, the charter and data discipline.