delegation-protocol
Skill0 starsChanged 3 months agoMITView on GitHub
How to route work across teams and agents in this repository. Use when a task spans more than one domain, when you are unsure which agent owns it, or before fanning out to several agents.
5 other projects use a version of this.
Adopted by
Most recently taken up by
- jeroromano/claude-code-orchestration-templatetheir own version, first seen Oct 14, 2026
- aws-solutions/konductortheir own version, first seen Sep 23, 2026
- gootf/software-agent-soptheir own version, first seen Aug 30, 2026
- megandmartin/agent-skills-repotheir own version, first seen Jul 24, 2026
- zebbern/termstacktheir own version, first seen Apr 16, 2026
Other versions5 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 5
- aws-solutions/konductor★ 12same name
- megandmartin/agent-skills-repo★ 4same name
- zebbern/termstack★ 4same name
- gootf/software-agent-sop★ 2same name
- jeroromano/claude-code-orchestration-template★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository16 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .claude/commands/agents.mdAGENTS.md
- .opencode/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- .cursor/rules/caveman.mdcCursor rule
- adrSkill
- agent-registrySkill
- cost-reviewSkill
- data-contractSkill
- marketplaceSkill
- onboardingSkill
- orchestrationSkill
- postmortemSkill
- runbookSkill
- threat-modelSkill
The file
We split this skill into the rules it gives an agent
2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 4 edited versions we read keep these.
2
Distinctive
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
- Distinctive: line 12
Never skip a link. Implementation with no design decision produces rework.
- Distinctive: line 20
Gather the declared Inputs: before delegating. An agent without its inputs guesses, and a confident guess is worse than an admitted gap.
“Common” is measured: kept by at least 6 in 10 of the 4 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Delegation Protocol
The chain
Architect decides. Engineer implements. Quality verifies. Release ships.
Never skip a link. Implementation with no design decision produces rework. A release with no verification produces an incident. Both cost more than the link you skipped.
Routing procedure
- Name the domain in one sentence. If you need two sentences, it is two tasks — split first.
- Find the owner in
.claude/teams/README.md. - Pick the minimum set. One agent per distinct concern. Not one per file, not one per idea.
- Gather the declared Inputs before delegating. An agent without its inputs guesses, and a confident guess is worse than an admitted gap.
- Delegate with context, not with the user's raw prompt. State what was already decided, what is out of scope, and what you want back.
- Reconcile. Where two agents disagree, do not average them. State the disagreement explicitly and escalate to the team's architect with both positions.
Fan-out budget
| Task | Agents | Why |
| One file, obvious fix | 0 — do it inline | Delegation costs more than the fix |
| One feature, one domain | 1-2 | Implementer, then reviewer |
| Cross-domain feature | 3-4, one per domain | Any more and they overlap |
| System design | Architects first, implementers in a second pass | Design before build, always |
More than four agents on one task means the task is not decomposed yet. Decompose it.
Worked example
"Our checkout times out under Black Friday load and we lose orders."
Wrong: spawn eight agents — backend, database, SRE, performance, QA, security, PM, release — and merge eight overlapping reports.
Right — sequential, each gated on the previous:
1. performance-test-engineer → where does it actually break?
Returns: "p99 8.2s at 400 rps; DB connection pool saturated at 200 rps, not CPU."
→ Everything below is now scoped by that finding. Without it, the rest is guesswork.
2. database-architect → pool sizing, connection strategy, read replica options
backend-architect → idempotency so retries do not double-charge
(parallel — different concerns, same evidence)
3. backend-developer → implements the agreed design
database-engineer → pool config and index changes
4. code-reviewer → correctness of the retry path
reliability-test-engineer → verify it degrades instead of collapsing
5. release-manager → canary with abort criteria tied to p99Five stages, nine agents total, but never more than two at once, and each stage gets real inputs from the last. Step 1 alone eliminated four wrong directions.
Anti-patterns
| Anti-pattern | What goes wrong |
| Spawning five agents to answer one question | Five overlapping reports, no decision, high cost |
| Delegating without declared inputs | Agent invents plausible numbers; you cannot tell |
| Skipping the architect on an irreversible decision | Rework at 10x the cost, three weeks later |
| Accepting output without checking evidence | You inherited a hallucination and now own it |
| Forwarding the user's raw prompt to the agent | Agent re-derives context you already have |
| Parallel agents on the same concern | They contradict; you arbitrate with no basis |
Done when
- The owning team is named and justified in one line.
- Each delegated agent received its declared Inputs.
- No two agents were given the same concern.
- Contradictions between agents are stated, not silently averaged.
- Every claim in the merged result traces to an agent's evidence, not to your summary of it.
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
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.

