I ran the most hyped AI agent catalog on GitHub through three subagents and installed exactly zero agents from it. This post is
The setup
The catalog: 273 agent files across 22 divisions, 169k stars, MIT license. Each file is a role with rules, deliverable templates and success metrics. Solid work, and mostly noise for any specific business: divisions for Douyin, WeChat, Unity, carousel generators for platforms you may not even be able to reach. The useful fraction is maybe 10 percent, but that fraction contains methodologies it'd take you months to polish yourself.
Our side: a CRM where an Obsidian vault is the single database and agents run the operations layer. Five scheduled roles (commercial director, finance director, CEO, marketer, on-call SRE), an hourly inbound cycle over email and messengers, and a company charter with hard gates. An agent physically cannot send an invoice below the price floor or promise a client a deadline without owner confirmation.
That context matters. A prompt is a personality without memory. What turns it into an employee's the bundle: skill plus schedule plus gates plus access to live company data.
The extraction run
Three subagents in parallel, each with a zone: marketing, sales and paid media (53 files), finance, PM, support, product and strategy (40 files), testing, engineering and specialized roles (about 70).
My instruction to each subagent: read everything, return a verdict per file in the context of our system. Steal, skip, or reference. Not a summary, a decision: what exactly to take and where exactly to put it.
By midday I had verdict tables for 160+ files, deep breakdowns of the top five per zone, and an implementation plan. Then the surgery: edits went into concrete role skills, and since our automation prompts reference those skills, every scheduled ritual picked up the new rules on its next run. Nothing was redeployed. Nothing restarted.
If you want to replicate this, the load-bearing detail we found is the verdict format. 'Interesting' is not a verdict. 'Steal the dunning ladder into the commercial director skill, replace our D+7 single reminder with a five step sequence' is.
What came out of it
The dunning ladder. Receivables now walk a fixed sequence: D-3 heads-up, D+1 reminder with the PDF, D+7 serious reminder with the acceptance act, D+14 formal demand plus a pause-work question, D+30 pre-legal. The first three steps send themselves.
Decisions with defaults. The behavioral design principle that tied it all together: every question to the owner arrives with a drafted answer. Not 'what do we do about this debtor' but 'letter drafted, send or edit'. One click beats an hour of deliberation, and default consent beats both.
The evidence protocol. This one targets the real disease of agent-run companies, fantasy reports. The rules we adopted:
- Every 'it works' claim requires proof: screenshot, curl output, log, green run. Otherwise the status is NOT TESTED.
- Four statuses: PASS, FAIL, NOT TESTED, NOT DETERMINED. NOT TESTED means the hole is in the testing, not a hidden all-clear.
- Automatic fail for 'no issues found' without a documented scope.
- Default verdict for fresh work: NEEDS WORK. First implementations take two or three revision cycles, that's normal.
- Perfect scores from a previous agent are a red flag. Reproduce two or three claims at random. The citation audit loop. Once a month, 20 to 40 live buyer prompts run through ChatGPT, Perplexity, Gemini and Yandex Neuro. The table records who gets cited. Lost prompts, the ones where a competitor wins, get a root cause check (comparison page, FAQ schema, freshness), a fix pack, and a re-check with the same prompt set 14 days later. Benchmark before fixes, always. Sales qualification. Six questions we now ask on every lead, no exceptions: what is broken, why, what does it cost, who else cares, why now, what happens if nothing is done. Plus the value equation as an offer auditor: value = (dream outcome x perceived likelihood) / (time delay x effort). Cut the price only after squeezing the other three levers. ## Why zero agents got installed Three reasons worth stating plainly:
- A personality without memory is not an employee. Those 273 prompts do not know our clients or our price floor.
- Idle agents burn attention and budget. Our roles wake in their slot, do the work, write the journal, sleep.
- Of the file format itself, only three sections survived contact with reality: hard rules (numbered Never/Always), a ready deliverable template, and verifiable metrics. Identity and vibe are cosmetics. ## Takeaways The competitive edge of 2026 isn't access to AI. Everyone has that. The edge is the speed of absorbing methodologies, because the world produces them faster than any human reads. A company with reading and adoption on a conveyor upgrades weekly. A company without one reads about upgrades on LinkedIn. Second: agents don't replace the owner, they prepare decisions. Approval mode instead of decision production. Third: honest systems beat pompous ones. We rejected dozens of prompts with '+300% leads' claims, numbers out of thin air. Verifiable rules went to production the same day. Apply the same filter to everything sold to you as a miracle. Originally published at https://automata.sale/en/blog/ai-automation/273-agenta-v-odin-den/