Retornatus

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).

ProjectCore betSimilar to Retornatus?
RetornatusRepo-native Change loop under .retornatus/; host writes code; Evidence ≠ Assurance—
Spec GuardrailsPredecessor; .specs/ + skill-driven phasesClosest lineage — same family, different architecture
TAUSIKFail-closed gates + ed25519 signed receiptsSimilar goals; heavier crypto/enforcement product
Claim PlaneControl plane around Codex: scope admission, diff verify, sealed digestsSimilar proof-of-delivery; more session/runtime control plane
MartinLoopBudgets, verifier gates, signed run records around CLIsSimilar gates; focuses on run budgets / wrappers
HighHarnessDefault-deny tool permissions + hash-chained episode logsSimilar audit trail; permission engine is host-layer (we stay native-first)
obsigna / Agent ReceiptsCryptographic receipt protocol (PyPI SDK + daemon)Complementary — we now emit a light HMAC receipt on verify; not a full receipt mesh
Microsoft AGTMCP-governed tool calls + offline receipts / SLSAEnterprise toolkit; different packaging surface
OpenAI harness practiceAGENTS.md map + mechanical CI invariantsPattern we adopt for discoverability — not a competing product
loop-harnessQualify/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.

What Retornatus uniquely emphasizes

  1. Situation as requirements analysis before Contract
  2. Evidence bound to Claims with Assurance as a separate judgment
  3. Complexity lanes (QUICK / STANDARD / COMPLEX)
  4. Human-controlled Skill intake (CREATE=yes)
  5. Native-first — prefer Cursor/Codex isolation over rebuilding sandboxes

What we borrowed as ideas (not code)

Idea in RetornatusInspired by (class)
verify --receipt (HMAC, local key)TAUSIK / obsigna / AGT receipt class
Optional Action attempt budgetMartinLoop budget class
AGENTS.md mapOpenAI harness practice
Docs HTML --check in CIDoc-gardening / mechanical invariants

No code was vendored from those repositories. Relationship is conceptual.

Honest overlap gaps

They often haveRetornatus stance
Live multi-provider agent runtime testsEnvironment owns agents (non-goal)
Enforced sandbox orchestrationAdvisory Boundaries + host native tools
Default-deny tool firewallHost policy; not rebuilt here
Full Ed25519 receipt mesh / SLSA exportStart with local HMAC receipt; expand if needed
← Docs hub Product site →