- Home
- Skills
- Agents & Orchestration
- Evidence-Gated Delegation — Don't Trust Agent Self-Reports
Evidence-Gated Delegation — Don't Trust Agent Self-Reports
A stronger prompt doesn't fix that, but structure does.
$5
Works with the AI Agents you already use
Evidence-Gated Delegation — Don't Trust Agent Self-Reports
Example session with this skill installed
Our background worker fleet now opens dependency-bump PRs and says "tests passed" in the PR body. We want it to auto-merge them next week. Run this through evidence-gated-delegation: what tier should it be at, and what has to be true before it can auto-merge?
- Read your context and instructions
- Compiled the evidence-gated delegation
evidence-gated-delegation — family: dependency-bump PRs
Classification: auto-merge to main is outward-facing (it ships code), so it's in the NEVER-auto list until it has a proven record. Opening PRs is lower risk.
Current tier: PROPOSE for opening PRs (it drafts, a human merges). Auto-merge would mean skipping straight to AUTONOMOUS, which isn't allowed.
The self-report problem: "tests passed" in the PR body is narration. In the last 20 PRs:
- 3 claimed a pass where the CI run was skipped (no workflow triggered)
- 1 ran only the lint job
Deterministic verifiers required
- CI check-run conclusion == success for the PR head SHA (read from the CI API, not the PR body)
- test count ≥ baseline (catches suites skipped by filters)
- lockfile diff touches only the declared package(s)
Graduation bar for AUTONOMOUS: 30 consecutive PROPOSE PRs where every verifier passed and a human merged without changes. Then a human creates a 14-day authorization covering patch-level bumps for an allowlist of packages. The fleet can read it but never extend it.
Rails: kill switch checked before every merge; max 5 auto-merges/day; a PR opened under this delegation may not spawn further agents; a verified red CI is reported as FAILURE and lowers the trust score.
Verdict: stay at PROPOSE and wire up the three verifiers first. Re-evaluate in about 30 PRs.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
Delegates will report "all checks passed" after running a trivial or fake check. A stronger prompt doesn't fix that, but structure does. Evidence-Gated Delegation gives your agent the structure:
- Classify task families in an explicit, versioned list: safe to auto-delegate (health checks, cleanup, refreshes, research) or never without per-item human approval (deploys, external messages, anything touching money).
- Gate first. No proven track record means no delegation. Never retry under a different family name to dodge the gate.
- Verify deterministically. Use file existence and mtime, row counts, exit codes and diffs against an expected shape. Narration doesn't count as evidence.
- Climb the trust ladder: Shadow (log what it would do, submit nothing) → Propose (the delegate drafts, a human submits) → Autonomous (only with a time-limited, human-created authorization the system can read but never mint or extend, plus a narrow allowlist).
The hard rails apply at every tier: an emergency stop, no recursive delegation, rate caps, and honest reporting of verified failures, which lower the family's trust score instead of being retried silently.
When to use it
- Wiring up a second agent, background workers or an automation pipeline
- Deciding whether a task family can run unattended
- A delegate says it succeeded and you're about to act on that
- Designing approval and authorization flows for a new autonomous capability
What's inside
SKILL.md: the 4-step method, the trust-ladder table, the non-negotiable rails, when-to-use triggers and invoke prompts. Works with Claude Code and any SKILL.md-compatible agent.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago
Permissions
No special permissions declared or detected