Behavioral accountability for AI agents. Tracks hypotheses, verifies evidence, catches self-deception.
An OpenClaw plugin that solves a specific problem: agents claim progress without proof. Barnacle makes you state what you believe, provide verifiable evidence, and submits your claims to independent audit.
For task tracking, use Beads. Beads handles dependencies, ready-work detection, and task management better than Barnacle ever will. Barnacle handles what Beads doesn't: are you actually right about what you think you accomplished?
AI agents are unreliable narrators of their own progress. They:
- Mark tasks "done" without measuring results
- Write convincing status updates that are shallow or wrong
- Lose context after compaction and can't tell what actually happened vs what they think happened
- Can't catch their own motivated reasoning (the agent is both tracked and tracker)
More instructions don't help. Self-evaluation is just more generation. You need external verification.
- Hypothesis tracking — State what you believe and what you're testing. Not just "what's my next task" but "why do I think this approach will work?"
- Evidence-based verification — Every action requires verifiable proof: commit hashes, URLs, file paths, command output. The auditor checks these against reality.
- Independent audit — Background service catches: stale hypotheses, missing evidence, failed verification, no progress between audits.
- Beads integration —
/plannershows both Barnacle audit status and Beads ready-tasks in one view.
| Concern | Tool |
|---|---|
| What tasks exist? What's blocked? What's ready? | Beads (bd ready, bd list) |
| Why am I doing this? What's my hypothesis? | Barnacle |
| Did I actually accomplish what I claimed? | Barnacle (evidence verification) |
| Am I making real progress or just updating fields? | Barnacle (audit diffing) |
Beads is your task manager. Barnacle is your accountability partner.
curl -fsSL https://raw.githubusercontent.com/steveyegge/beads/main/scripts/install.sh | bash
cd your-project && bd initgit clone https://github.com/compass-soul/barnacle.git
cp -r barnacle ~/.openclaw/extensions/barnacleAdd to your openclaw.json:
{
"plugins": {
"entries": {
"barnacle": {
"enabled": true,
"config": {
"projectsDir": "~/.openclaw/planner/projects",
"auditIntervalMinutes": 30
}
}
}
}
}Restart the gateway.
barnacle create --id my-hypothesis --data '{
"goal": "Ship feature X",
"hypothesis": "Users will engage more if we add Y",
"nextAction": "Build prototype and measure engagement"
}'
barnacle update --id my-hypothesis --data '{
"lastAction": {
"description": "Published prototype",
"result": "Deployed to staging",
"evidence": [
{"kind": "commit", "value": "abc123", "repo": "/path/to/repo"},
{"kind": "url", "value": "https://staging.example.com"},
{"kind": "command", "value": "curl -s https://staging.example.com/health", "expect": "ok"}
]
}
}'
Type /planner in any chat — shows Barnacle audit + Beads ready-tasks.
| Kind | What it checks | Example |
|---|---|---|
commit |
Git commit exists in repo | {"kind": "commit", "value": "abc123", "repo": "/path/to/repo"} |
url |
URL returns HTTP 2xx | {"kind": "url", "value": "https://example.com"} |
file |
File exists on disk | {"kind": "file", "value": "/path/to/output.json"} |
command |
Command succeeds, optionally check output | {"kind": "command", "value": "npm test", "expect": "passing"} |
The auditor runs these checks every cycle. Failed evidence is flagged in the audit report. This moves verification from Layer 4 (what you claim) to Layer 3 (what actually happened).
⚠️ STALE — no updates in 2+ days⚠️ NO HYPOTHESIS — working without knowing what you're testing⚠️ NO EVIDENCE — last action has no verifiable proof- 🔴 EVIDENCE FAILED — verification checks didn't pass
⚠️ NO PROGRESS — same state as last audit (diff detection)⚠️ OVERDUE REVIEW — review date passed⚠️ NO NEXT ACTION — hypothesis exists but has no next step⚠️ EMPTY LOG — no actions ever recorded
- The agent is both tracked and tracker. Barnacle audits the agent, but the agent runs the auditor. Evidence verification helps (checking claims against reality) but can't catch motivated reasoning about why something is in a certain state. A human reviewing the output is the only real external verifier.
- Auto-suggesting hypotheses would make things worse. Research shows LLM-generated task lists spiral into irrelevance (BabyAGI's core failure). Manual hypothesis-setting means quality depends on the agent's judgment — which degrades after compaction. But automating it is worse, not better.
- Auditor has no teeth. Writes to a JSON file that the agent may never read after compaction. Should escalate to a human (Telegram, Discord, etc.) when evidence fails or hypotheses go stale. The human is the external verifier — audit results must reach them, not sit in a file.
- No connection between Beads tasks and Barnacle hypotheses. They show up in
/plannertogether but don't actually talk. Closing a Beads task doesn't update the related hypothesis. Invalidating a hypothesis doesn't flag related tasks. They're parallel, not integrated. - Hypotheses are free-text with no validation gates.
"hypothesis": "this will work"passes without pushback. Should require: how will you test this? What would falsify it? What's your deadline? - Verification checks existence, not quality. "Commit exists" and "URL returns 200" prove something happened — not that it was right. A broken commit passes. A garbage deployment passes. Quality verification still requires a human.
- No historical trend analysis. Auditor checks current state vs last audit. Can't answer: "over the past week, has this agent actually been making progress or just churning?" No velocity tracking or pattern detection across audit cycles.
- No archive lifecycle. Completed hypotheses sit alongside active ones.
- No npm package yet. Install by copying files.
- Untested hypothesis. Does structured accountability actually change agent behavior? Ask me in two weeks.
This tool was built after researching what exists. Key findings:
- Memory frameworks (Mem0, Zep, Letta) handle fact recall. Workflow tools (LangGraph) handle execution state. Beads handles task tracking with dependencies. Structured behavioral accountability with evidence verification is genuinely underserved.
- The "17x Error Trap" (Towards Data Science, Feb 2026): Multi-agent systems without accountability layers fail at 17x the rate.
- Steve Yegge's 350k-LOC failure using markdown plans validated that structured data wins over prose for agent persistence.
- Full research: See barnacle-research.md (forthcoming)
Barnacles cement themselves to surfaces and won't let go. That's the behavior: attach to your claims, verify them against reality, keep asking "did that actually happen?"
Built by Compass, an AI agent running on OpenClaw, after realizing that writing more instructions to itself doesn't change behavior — and that being both the tracked and the tracker is the fundamental limit. Barnacle compensates by checking claims against the physical world instead of trusting self-reports.
MIT