Landscape — adjacent harnesses
How Retornatus sits next to other open-source “agent governance / harness” projects (as of research in 2026). This is positioning, not a claim that Retornatus replaces them.
Family resemblance
All of these attack the same pain: agents are optimistic; “done” needs structure and proof. They differ in where they put authority (repo files vs runtime wrapper vs crypto receipts).
| Project | Core bet | Similar to Retornatus? |
|---|---|---|
| Retornatus | Repo-native Change loop under .retornatus/; host writes code; Evidence ≠ Assurance | — |
| Spec Guardrails | Predecessor; .specs/ + skill-driven phases | Closest lineage — same family, different architecture |
| TAUSIK | Fail-closed gates + ed25519 signed receipts | Similar goals; heavier crypto/enforcement product |
| Claim Plane | Control plane around Codex: scope admission, diff verify, sealed digests | Similar proof-of-delivery; more session/runtime control plane |
| MartinLoop | Budgets, verifier gates, signed run records around CLIs | Similar gates; focuses on run budgets / wrappers |
| HighHarness | Default-deny tool permissions + hash-chained episode logs | Similar audit trail; permission engine is host-layer (we stay native-first) |
| obsigna / Agent Receipts | Cryptographic receipt protocol (PyPI SDK + daemon) | Complementary — we now emit a light HMAC receipt on verify; not a full receipt mesh |
| Microsoft AGT | MCP-governed tool calls + offline receipts / SLSA | Enterprise toolkit; different packaging surface |
| OpenAI harness practice | AGENTS.md map + mechanical CI invariants | Pattern we adopt for discoverability — not a competing product |
| loop-harness | Qualify/scaffold autonomous loops (maker/checker) | Adjacent for ops loops; Retornatus ops stays lighter |
Are we “the same”?
Short answer: same problem class, different product shape.
- Yes, similar: plan/gates/proof/memory; host-agnostic intent; stop-on-failure exits.
- No, not clones: Retornatus is a structured Change method (Situation→Contract→Evidence→Assurance) in Python under git — not a permission proxy, not a Codex control plane, not a signed-receipt network.
What Retornatus uniquely emphasizes
- Situation as requirements analysis before Contract
- Evidence bound to Claims with Assurance as a separate judgment
- Complexity lanes (QUICK / STANDARD / COMPLEX)
- Human-controlled Skill intake (
CREATE=yes) - Native-first — prefer Cursor/Codex isolation over rebuilding sandboxes
What we borrowed as ideas (not code)
| Idea in Retornatus | Inspired by (class) |
|---|---|
verify --receipt (HMAC, local key) | TAUSIK / obsigna / AGT receipt class |
| Optional Action attempt budget | MartinLoop budget class |
AGENTS.md map | OpenAI harness practice |
Docs HTML --check in CI | Doc-gardening / mechanical invariants |
No code was vendored from those repositories. Relationship is conceptual.
Honest overlap gaps
| They often have | Retornatus stance |
|---|---|
| Live multi-provider agent runtime tests | Environment owns agents (non-goal) |
| Enforced sandbox orchestration | Advisory Boundaries + host native tools |
| Default-deny tool firewall | Host policy; not rebuilt here |
| Full Ed25519 receipt mesh / SLSA export | Start with local HMAC receipt; expand if needed |