311 Agents, No Fleet

§1The find

A repository called agency-agents crossed 145,000 GitHub stars this week, and the pitch is hard to argue with: a complete AI agency in a folder. Install it and you get 311 ready-made agents, from backend architects to brand guardians to a "whimsy injector", each one a specialist with a personality and a checklist. It is MIT-licensed, it installs into Claude Code, Cursor, Codex and Gemini with one command, and people are voting for it with their stars.

We run a multi-agent fleet ourselves, so this is exactly the kind of thing we should either borrow from or have an opinion about. We pulled it apart. The short version: the repo is good, and the number on the tin is the least interesting thing about it.

§2What is actually in there

311 agent files, spread across 21 divisions. Engineering is the biggest at 58, then a "specialized" catch-all at 57, marketing at 36, game development at 21, and a long tail down through security, finance and healthcare. Each agent is a single markdown file with the same clean shape every time: a frontmatter block (name, one-line description, a colour, an emoji, a "vibe") and then a system prompt with a fixed skeleton, identity, core mission, hard rules, and a checklist.

The quality is real. The code-reviewer agent, to pick one, tells the model to grade findings as blocker, suggestion or nit, to explain the why rather than just the what, and to deliver the whole review in one pass instead of drip-feeding comments across rounds. That is a genuinely sensible review prompt. Multiply it by 311 and you have a tidy, opinionated prompt library that someone has clearly sweated over.

§3What it cannot do

Here is the catch, and it is the whole point. Every one of those 311 agents is a standalone prompt. You install them into your coding tool and invoke one by name. None of them can reach another. There is no routing that decides which agent should take a task. There is no shared memory, so nothing one agent learns is available to the next. There is no escalation path when an agent is out of its depth, and no gate that checks an agent's output before it counts as done.

That is not a knock on the repo. It is honest about what it is: a roster you install, not a system you run. The orchestration is whatever your host tool happens to do, which today is mostly "the human picks the agent and reads the result". The 311 files are the cast list. There is no script, no director, and no stage.

§4The roles were never the hard part

This is the bit worth internalising if you are building with more than one agent. Writing a good role prompt is the easy 20%. We know, because we have written a fair few, and it is an afternoon each. The part that takes the real time, and the part that decides whether a fleet works, is everything the markdown file cannot hold:

• Routing. Something has to decide which agent gets the job, and hand off cleanly when the wrong one picked it up. Get this wrong and your clever specialists never fire.
• Shared memory. An agent that cannot see what the last one already found repeats the work, or worse, contradicts it. The value compounds only when the memory is shared.
• Escalation. A real fleet knows when an agent is stuck and pulls in a bigger model or a human, instead of confidently returning nonsense.
• A gate. The highest-leverage piece by a distance. Something scores the output against a bar before it ships, so a weak answer gets caught and revised rather than shipped. We have measured this: the same model, on the same task, with the same prompt, scored 51 and 97 with and without a gate. The gate did the work, not the prompt.

None of that lives in a role file. All of it lives in the wiring between the files. A library of 311 prompts gives you none of it, and a library of 3,110 would give you exactly as much.

§5So is it useful?

Yes, with the ceiling drawn honestly. As a prompt library it is a good one: the per-agent structure (frontmatter plus a fixed identity, mission, rules and checklist skeleton) is worth copying wholesale, and there are maybe five to ten role prompts in there sharper than what you would write cold. Take the structure, take a handful of the roles, and expect to rewrite even those to fit how you actually work.

As a team, it is not one, and no catalogue can be. The star count measures how many people want a shortcut to good role prompts, which is a real and reasonable want. It does not measure orchestration, because orchestration is the part nobody can hand you in a folder. If you are deciding where to spend your next week of agent-building, spend it on the wiring. The roles you can borrow. The fleet you have to build.