Continuity for small teams whose agents hand work to each other.
When an agent stops, times out, or changes providers, Trust Plane preserves the task, revalidates its authority, prevents duplicate action, and requires evidence before the resumed work counts as complete.
The promise
Connect your existing agents once. The Trust Plane keeps their work coordinated, recoverable, approved, and verifiably complete, without replacing the model host or framework you already use.
It does not claim to verify your production agents, your security posture, or your business results by itself. It verifies the requests, approvals, handoffs, and evidence that pass through it, and it says so plainly when something has not been checked.
The governed request-to-receipt workflow
Every skill call takes the same bounded path, whether it arrives over A2A, MCP, or REST:
Register
An agent and its skills are recorded in the private registry with an owning team, environment, and sensitivity ceiling. Each skill declares its effect, whether it needs approval, and the exact input it accepts.
Request
A caller sends a bounded request. The gateway will only route to a destination already in the registry, and it rejects input that does not match the declared schema.
Approval binding
When a skill requires approval, the exact operation and arguments are bound to that approval. A different authenticated person must resolve it, and an approval left over from a stale or expired lease is rejected, not reused.
Execution
The gateway reserves the operation, applies an idempotency key so a retry cannot duplicate an effect, and enforces time and response-size limits on the call.
Evidence
The result is reduced to bounded evidence and a digest, not a raw payload. Where independent verification is required, a separate principal checks the outcome against the original objective before it can be marked complete.
Receipt
The full chain is written to an append-only, digest-chained event log. /v1/audit/verify confirms the chain has not been altered.
These are synthetic proof scenarios shaped around portfolio product workflows. They use fixture or public data only, and each one states plainly what it did and did not verify.
Verified synthetic runtime
CounselorAI: cross-agent continuity
A CounselorAI-shaped task moves between simulated Codex and Claude principals after a lease expires. The controlled-clock proof records a simulated 1.0 second recovery interval, rejects the stale approval, binds a fresh approval to the new lease, prevents duplicate execution, and requires independent verification before completion.
Result: 30 of 30 runtime points. It uses no real client data and makes no call to CounselorAI. The downloadable proof includes synthetic scenario context, event digests, and measurements. This is not a measurement of live model recovery.
Verified synthetic runtime
Breadcrumbs: governed memory continuity
A minimized memory capsule survives an interruption between simulated Codex and Claude principals. The recovered agent receives a fresh lease and approval, supersedes stale memory with a correction, and completes only after independent verification. The controlled-clock proof records a simulated 1.0 second recovery interval.
Result: 30 of 30 runtime points. It also proves private-field minimization, public PII rejection, stale-memory invalidation, and a valid event chain. It uses synthetic records and does not read or modify the Breadcrumbs repository or invoke either model.
Evidence boundary: both scenarios prove that the Trust Plane's controls behave correctly against a realistic workflow shape. Neither one is a claim about CounselorAI's or Breadcrumbs' production security, uptime, or business outcomes.