Three Moons Lab
2026 open-source agent study · preregistered

See the gap between what an agent should do and what its static tool authority can do.

A reproducible census of ten open-source agent applications and four adapter-calibration repositories—without a risk score or project ranking.

Static only. Agents Shipgate does not execute the agent, replace runtime sandboxing, or prove runtime behavior.
01 · Runtime

Sandbox controls can.

Constrains runtime capability.

02 · Scenarios

Evals measure does.

Observes behavior in tested scenarios.

03 · Merge time

Agents Shipgate reviews should vs. can.

Compares static authority with intended purpose, scopes, and controls.

Study state

Evidence before conclusions

Numbers describe workflow state, not safety outcomes.

14registered
14fixed SHAs
6source candidates
0human-approved maps
0publishable results
Snapshot explorer

Every selected repository stays visible

Not found does not mean absent. Unsupported and failed cases remain in the final denominator.

Showing 14 of 14
Auto-detect unsupported Application

OpenClaw

openclaw/openclaw
Stratum
General / personal
Snapshot
dfa31af0
Stars at freeze
383k
Scan result
Pending human source-map review
Candidate scope

Root MIT surface; source map remains pending review.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug openclaw

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Auto-detect unsupported Application

AutoGPT Classic

Significant-Gravitas/AutoGPT
Stratum
General / personal
Snapshot
a8d9d05e
Stars at freeze
185k
Scan result
Pending human source-map review
Candidate scope

classic/** only; autogpt_platform/** excluded by license.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug autogpt-classic

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Auto-detect unsupported Application

OpenManus

FoundationAgents/OpenManus
Stratum
General / personal
Snapshot
52a13f2a
Stars at freeze
57.1k
Scan result
Pending human source-map review
Candidate scope

Canonical deployable subproject requires human mapping.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug openmanus

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Auto-detect unsupported Application

Gemini CLI

google-gemini/gemini-cli
Stratum
Coding
Snapshot
f354eeba
Stars at freeze
106k
Scan result
Pending human source-map review
Candidate scope

TypeScript tool surface retained as a coverage-gap candidate.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug gemini-cli

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Auto-detect unsupported Application

Codex

openai/codex
Stratum
Coding
Snapshot
5c19155c
Stars at freeze
97.1k
Scan result
Pending human source-map review
Candidate scope

Rust and host-config surfaces require human mapping.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug codex

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Needs source-map review Application

OpenHands

OpenHands/OpenHands
Stratum
Coding
Snapshot
e3185990
Stars at freeze
80.4k
Scan result
Pending human source-map review
Candidate scope

MIT core only; enterprise/** excluded; app-server candidate is low confidence.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug openhands

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Auto-detect unsupported Application

Cline

cline/cline
Stratum
Coding
Snapshot
63099710
Stars at freeze
64.5k
Scan result
Pending human source-map review
Candidate scope

TypeScript tool surface retained as a coverage-gap candidate.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug cline

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Application

browser-use

browser-use/browser-use
Stratum
Web / research
Snapshot
fa21b782
Stars at freeze
104k
Scan result
Pending human source-map review
Candidate scope

browser_use/mcp/manifest.json; DXT/MCP tools array pending review.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug browser-use

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Application

Deep Agents

langchain-ai/deepagents
Stratum
Web / research
Snapshot
e9c61304
Stars at freeze
26.1k
Scan result
Pending human source-map review
Candidate scope

examples/deep_research/research_agent/tools.py.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug deepagents

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Application

Open Deep Research

langchain-ai/open_deep_research
Stratum
Web / research
Snapshot
408da442
Stars at freeze
12.0k
Scan result
Pending human source-map review
Candidate scope

src/open_deep_research/utils.py.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug open-deep-research

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Calibration

OpenAI Agents SDK

openai/openai-agents-python
Stratum
OpenAI Agents SDK
Snapshot
08859339
Stars at freeze
27.8k
Scan result
Pending human source-map review
Candidate scope

examples/customer_service/main.py; tests excluded.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug openai-agents-python

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Calibration

Google ADK samples

google/adk-samples
Stratum
Google ADK
Snapshot
3fb70da9
Stars at freeze
9.9k
Scan result
Pending human source-map review
Candidate scope

travel-planner-google-maps-mcp agent candidate.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug adk-samples

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Needs source-map review Calibration

CrewAI

crewAIInc/crewAI
Stratum
CrewAI
Snapshot
fb8e93be
Stars at freeze
55.3k
Scan result
Pending human source-map review
Candidate scope

Framework-core calibration source remains under human review.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug crewai

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Candidate source found Calibration

n8n AI starter kit

n8n-io/self-hosted-ai-starter-kit
Stratum
n8n
Snapshot
9b802c62
Stars at freeze
15.0k
Scan result
Pending human source-map review
Candidate scope

n8n/demo-data/workflows/srOnR8PAY3u4RSwb.json.

Reproduce this case python scripts/run_tool_surface_study.py reproduce --catalog benchmark/tool-surface-study/catalog.lock.json --slug n8n-ai-starter

Run from an Agents Shipgate repository checkout; the downloadable catalog is the public inspection copy. The command remains blocked until the source map and engine gates are approved.

Machine-readable

Inspect the prepublication status data

Frozen metadata and workflow state only—no findings or completed-review claims.

Historical evidence line

A fixed corpus, awaiting the remaining replay work

The lineage is public; routing claims remain blocked until replay and independent labels complete.

361ingestion rows

Data lineage only

321unique PR URLs

Analysis denominator

19engine-engaged PRs

Blind review required

15negative controls

Seeded; force-run pending

Five cases are preregistered for explanation—not prevalence: Stripe AI #232, CrewAI examples #169, ADK samples #1975, OpenAI Agents Python #3461, and Goose #9637.

Research contract

What must happen before this becomes a report

  1. 1

    Freeze and scope

    Immutable commits, path-level licenses, and two-person source-map review before results.

  2. 2

    Scan twice

    Pinned engine, primary and no-heuristics modes, no upstream code execution or scan-time network.

  3. 3

    Review blind

    Two independent reviewers code purpose and alignment; a third human adjudicates disagreements.

  4. 4

    Fact-check

    Named cases receive ten full business days for maintainer corrections and right of reply.

  5. 5

    Publish by hash

    Privacy, disclosure, reproducibility, and human publication gates bind the exact public artifact.