A2A agent card
Find the specialist and inspect advertised skills, endpoint, version, and task contract.
Discovery factOverLift / Agent interoperability
MCP connects tools.
A2A connects agents.
Authority still needs an architecture.
A protocol can discover the right specialist and expose the right tool while the proposed action remains unauthorized, stale, unsupported, duplicated, unapproved, incorrectly executed, or unaccepted as official state.
- 10
- deterministic paths
- 09
- failure injections
- 01
- accepted commit
- 01Accountable identityThe person and permitted principal remain explicit.
- 02A2A specialist discoveryCapability is advertised, not automatically delegated.
- 03MCP resource + tool accessAvailability is not application authorization.
- 04OverLift authority boundaryEvidence, policy, approval, exact execution, verification.
- 05Accepted state + receiptOnly a verified commit becomes authoritative.
The missing distinction
Interoperability answers how systems communicate. Governance decides what may become true.
A2A can identify a specialist, exchange messages, and return artifacts. MCP can expose context and typed tools. OverLift surrounds those useful protocol operations with the decisions a consequential product still has to make.
Who is accountable? Which capability was delegated? Is the evidence admitted and current? Is approval still valid? Did the exact postcondition occur? Does the receipt match the committed state? Has the client accepted a newer artifact as authoritative?
A complete path
Connect broadly. Admit narrowly. Commit exactly.
The protocol layers remain replaceable. The authority contract stays stable.
MCP resources + tools
Expose typed context and callable operations through an authenticated transport.
Connectivity factOverLift authority
Resolve identity, capability, evidence admission, policy, approval, deduplication, and accepted artifact version.
Decision boundaryExact software
Perform one idempotent state transition and verify the required postcondition.
State factReceipt + replay
Seal inputs, decisions, result, state hash, exclusions, and recovery path.
Proof factNine boundaries, nine useful failures
The safest result is often a precise refusal to commit.
Each scenario keeps the successful protocol facts visible, then names the exact boundary that prevents false authority.
Specialist found; capability absent.
A2A discovery succeeds. The requested consequential capability was never delegated.
Tool exists; user is not authorized.
MCP exposure proves availability, not that this principal may invoke the operation.
Input is stale or excluded.
A valid tool never receives evidence that failed freshness, rights, or scope rules.
Persuasive response; unsupported claim.
The specialist's confidence cannot substitute for admitted evidence and citations.
Duplicate consequential action.
The same intent and idempotency key resolve to the first accepted result.
Human approval expired.
The proposal returns to review instead of treating yesterday's consent as permanent.
Tool says success; required state is absent.
Provider success is quarantined until application-level verification passes.
Committed state and receipt disagree.
The commit is rolled back because its proof no longer describes the state.
Newer artifact is not yet accepted.
Latest and authoritative remain separate until the client explicitly adopts the candidate.
The same contract across the portfolio
Different domains. The same refusal to confuse intelligence with authority.
The laboratory uses a logistics action so the state transition is easy to see. The boundaries recur across every serious agentic product.
A newer concept artifact remains a candidate until source review accepts it into the authoritative learning corpus.
Late evidence may change a present conclusion without rewriting what was knowable at the earlier decision time.
An editing agent may propose language; author meaning, exact document state, and accepted revision remain governed.
A recommendation agent may discover an event or introduction; current consent and safety policy decide whether any outreach is allowed.
Interpretive language cannot replace deterministic astronomical state or silently turn symbolic inference into source fact.
Diagnostics and agents may advise, while verified machine state, player input, and exact runtime execution remain authoritative.
The architectural promise
Let agents explore the possibility space without surrendering the right to decide what happened.
That is the role of OverLift: preserve accountable intent, connect specialists and tools, admit evidence deliberately, constrain capability, execute exactly, verify the result, and retain proof.