OpenAI Agents SDK
Wrap a tool before the provider call, surface an authority interruption, then resume within the approved grant.
@gomission/openai-agents-adapterOPEN AUTHORITY LAYEROPENAI AGENTS SDKMCP 2025-11-25
Mission Gate
Agents earn permission to do real work.
Mission governs the moment an AI agent crosses from prepared work into real-world consequence.
Exact action. Exact target. Exact input. Narrow grant. Human decision. Bounded execution. Signed receipt.
The missing primitive
Approval is a UI event. Authority is a verifiable contract.
Agent frameworks can pause for a person. Mission binds that decision to the exact action that may execute—and proves afterward that nothing else did.
Before executionTyped action envelope, consequence classification, exact payload and target digest.
At the boundaryTenant-bound, expiring, revocable, single-use or bounded-sequence authority.
After executionEd25519-signed receipt correlated to decision, grant, provider result, and OpenTelemetry trace.
When anything changesReplay, input mutation, target mutation, expiry, revocation, and tenant mismatch fail closed.
Reference implementation
One authority vocabulary across agent runtimes.
Mission adds the authority boundary without replacing your agent framework, IAM, policy engine, provider, or observability stack.
MCP 2025-11-25
Govern consequential MCP tool calls with version negotiation, fail-closed lifecycle handling, and explicit legacy exceptions.
@gomission/mcpCanonical schemas
Action, grant, decision, execution, receipt, signature, and public-key contracts with a conformance CLI.
@gomission/mission-schemasSigned receipts + OTel
Cryptographically bind policy and payload digests to trace context and provider outcomes for offline verification.
mission-execution-receipt/v2Seven gates
The happy path is not the proof. Failure behavior is.
Mission is built around the cases that ordinary approval callbacks leave ambiguous.
Replay protectionTarget integrityInput integrityTenant isolationExpiryRevocationBounded scope
Trust Graduation
Cross-runtime earned authority.
Trust Graduation is Mission’s method and open protocol profile: agents start with preparation rights and earn bounded execution rights from evidence—not model confidence.
Mission Control is the reference approval interface and dogfood environment. Mission Gate is the portable infrastructure product.
Research, draft, rank, and stage with no external consequence.
Classify the effect and present the exact boundary to a person.
Run only the approved target and payload within time and scope.
Receipts and outcomes influence future authority; preferences never authorize execution.
Three launch partners
Put one consequential action behind a production-grade authority boundary.
A focused 4–6 week production pilot for AI platforms and agent products already approaching real external actions.
01Select one runtime and action class
02Integrate the 30-minute governed-action proof
03Run replay, mutation, expiry, revocation, and tenant tests
04Verify signed receipts outside Mission
05Make a production, extend, or stop decision
The 30-minute proof
Attempt → interrupt → approve → execute within scope → verify.
If the action changes after approval, Mission refuses to execute it.