agentleFS
impactbrusselsimpactbrussels/AINativeOS/skills/gateway-agent-ops/SKILL.md

gateway-agent-ops

Skill1 starsChanged 5 months agoApache-2.0View on GitHub

Use when one agent is trying to do every operational job and it is drifting, getting confused, or running up a bill - when the founder says "my agent does everything", "it routed a refund through the wrong logic", "it got stuck in a loop all weekend", "the prompt is 4000 words and nobody can touch it", "I got a four-figure API bill", "should I build one big assistant for ops", or wants to automate the boring operational tail safely. Produces a routing map (request classes to specialist agents), a classification schema, and the loop and cost guardrails. Run after write-the-claude-md; pairs with self-healing-fallbacks.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Write down every kind of request the inbox actually receives, then collapse them to the smallest set a narrow specialist could own.

    line 19
  2. Distinctive:

    No specialist ships without all four.

    line 42
  3. Distinctive:

    No specialist ships without a budget cap and an iteration limit.

    line 66

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Gateway Agent Ops

One big assistant for operations is the trap that wears a different coat. For a week it is magic, then it routes a refund through the marketing logic, books a board call for 3am, and burns tokens all weekend reconciling two records that cannot be reconciled. Monday brings a four-figure bill and a system nobody trusts. The fix is not a better mega-prompt. It is a thin switchboard in front of a roster of narrow specialists, each one watched, capped, and easy to fix when it misbehaves.

The method

Build the switchboard, not the hero. Full framework: references/gateway-design.md. Source: Handbook Chapter 13.

Step 1: List the request classes

Write down every kind of request the inbox actually receives, then collapse them to the smallest set a narrow specialist could own. One class, one job.

Request classOne-job specialistSmall toolset
Summarise this threadSummariserread-only
Draft this grant sectionGrant-drafterretrieval + template
Check this against the rulebookCompliance-checkerrulebook lookup
Find a slot and book itSchedulercalendar write

Step 2: Design the classification schema

The gateway reads the request and emits one structured verdict: {class, specialist, confidence, reason}. It does one job, route, and gets out of the way. The moment it carries domain logic you have rebuilt the God Agent with an innocent name.

Step 3: Set the route / escalate / refuse rule

For each verdict, decide the disposition before going live.

SituationDisposition
Confident match to a specialistRoute
Low confidence, or no specialist fitsEscalate to a human
Irreversible action (filing, payment, send)Route to draft, human signs
Out of scope or unsafeRefuse, log, surface

Step 4: Wrap every specialist in guardrails

No specialist ships without all four. These are the seatbelt, not the polish.

GuardrailWhat it stops
Max-iteration capThe agent retrying the same failed step forever
Budget ceiling (per task)The meter spinning while you sleep
Wall-clock timeoutA subtly impossible goal running all weekend
Loop detection"I have tried this exact thing three times"
Human escalation pointAn irreversible call made without a signature

Step 5: Aim the machine at the boring tail

Rank candidate tasks by total time eaten, not by how annoying they feel. Build the specialist for the forty-minutes-a-day leak nobody complains about, not the loudest chore.

Output

  • A routing map: request classes mapped to one-job specialists, each with its small toolset.
  • A classification schema: the structured verdict the gateway emits, plus the route / escalate / refuse rule.
  • The guardrail set per specialist: max steps, budget cap, timeout, loop detection, human escalation point.
  • Next: run self-healing-fallbacks to give each specialist a bounded recovery and a deterministic safe answer.

Constraints

  • Name and forbid the God Agent: if you are adding a business rule to the router, that rule belongs in a specialist, not the gateway. Keep the gateway dumb.
  • Apply human-on-the-loop: fix the escalation point and the deterministic stop before the specialist goes live, not after the first close call. Every irreversible action drafts; a human signs.
  • No specialist ships without a budget cap and an iteration limit. An agent that can spend money in a loop without a ceiling is a liability you have not been billed for yet.
  • Stay theme-agnostic; the founder supplies the domain, you supply the rigour.

Dictionary

gateway agent · the God Agent · agent · infinite loop (runaway agent) · human-on-the-loop

Copy-paste version

For non-coders: paste into any chatbot or no-code builder (Claude.ai, ChatGPT, Lovable, Bolt, v0).

Act as my operations systems engineer. I want to automate my boring operational tail without one
big assistant that drifts and runs up a bill. Domain: [DOMAIN]. The work I want handled: [PROBLEM].
1. List the request classes my ops inbox actually receives, then collapse them to the smallest set
   of one-job specialists. For each: name it, give it ONE job and a small toolset.
2. Design the gateway's classification schema: it reads each request and outputs only
   {class, specialist, confidence, reason}. It routes and does nothing else. Tell me off if I try
   to put domain logic in the router - that is the God Agent with a new name.
3. Give me the route / escalate / refuse rule: when to route to a specialist, when to escalate to
   me, when an irreversible action [GUARDRAIL] must draft and wait for my signature, when to refuse.
4. For every specialist, write the guardrails it must ship with: a max-iteration cap, a per-task
   budget ceiling, a wall-clock timeout, loop detection, and the point where I sign before anything
   irreversible. No specialist ships without all of these.
5. Rank my candidate tasks by total time eaten per week, not by how annoying they are, and tell me
   which one specialist to build first.
Do not flatter the design. End with the single next action: build the first specialist with its
guardrails, behind the gateway.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.