Applied OverLift / Field note 01
MCP connects tools.
A2A connects agents.
Neither owns authority.
Interoperability is necessary infrastructure for agentic systems. It is not the final answer to identity, delegated power, admitted evidence, human approval, exact execution, official state, or recovery.
A
n agent finds a specialist. The specialist sees a useful tool. Authentication succeeds. The invocation returns success. None of those facts, alone or together, proves that the user was allowed to request the action, that the evidence was current, that approval remained valid, that the intended postcondition occurred, or that the resulting artifact became authoritative.
This is not a criticism of MCP or A2A. It is a boundary problem. Protocols become safer when a product stops asking them to own decisions outside their contract.
01 / What the protocols solve
Connectivity and collaboration are real architectural victories.
MCP gives an AI application a typed way to discover context and invoke tools. A2A gives separately built agents a way to advertise skills, exchange messages, manage tasks, and return artifacts. Together they reduce bespoke integration work and make specialist ecosystems practical.
Resources, tools, prompts, transport, and an integration contract.
The host can enumerate what exists and call a structured operation without teaching every model a private API dialect.
Discovery, messages, tasks, status, and artifacts across agent boundaries.
An orchestrator can find a specialist and collaborate without requiring access to that specialist's internals.
The mistake is not using these protocols. The mistake is turning protocol success into a universal green light.
02 / What remains
The product still has to decide what may become true.
Can this system communicate? is a different question from: May this principal cause this exact state transition using this evidence right now?
A consequential application still needs an accountable identity, a bounded delegation, application-level authorization, evidence admission, freshness and exclusion rules, a typed proposal, policy evaluation, meaningful approval, idempotency, exact execution, postcondition verification, an accepted artifact version, and a receipt that can reproduce or reverse the result.
OverLift treats those as one governed state machine. Learned components may search, recommend, explain, compare, and abstain. Protocol layers may transport resources, calls, messages, and artifacts. Exact software owns the transition into official state.
03 / Nine boundary failures
A mature demo should prove refusals, not only the happy path.
- Delegation: the specialist is discoverable but lacks the requested delegated capability.
- Authorization: the MCP tool exists, while the accountable user lacks permission to invoke it.
- Evidence admission: a valid operation is proposed from stale, excluded, unlicensed, or out-of-scope evidence.
- Support: the A2A response is articulate and confident but cannot point to admitted support.
- Idempotency: a retry attempts the same consequential action a second time.
- Approval: a previously valid human decision expires before execution begins.
- Postcondition: the provider reports success while the required application state is absent.
- Receipt integrity: the committed state hash no longer matches the sealed receipt.
- Artifact authority: a specialist returns something newer, but the client has not accepted it as the official version.
Each failure needs a different result. Some actions are blocked before execution. Some are held for fresh evidence or approval. Some are deduplicated. Some are quarantined after a provider response. Some require rollback. A newer artifact may remain visible as a candidate while the prior accepted artifact stays authoritative.
04 / The OverLift path
Connect broadly. Admit narrowly. Commit exactly.
accountable request
↓
A2A specialist discovery
↓
MCP context and typed tools
↓
identity + delegated capability
↓
evidence admission + support
↓
policy + approval + deduplication
↓
exact state transition
↓
postcondition verification
↓
receipt + accepted artifact + replay
The orchestration layer may change models, agents, or providers without changing this contract. That separation is valuable. Protocol innovation can move quickly while authority remains testable and deterministic.
05 / Product consequences
The boundary changes shape, but it does not disappear.
In OverLift Logistics, the consequential state may be a shipment reroute. In Vellucent, it is the accepted document revision and the author's preserved meaning. In Veluris, it is whether a recommendation may become outreach under present consent. In Saros, it is the conclusion supported by evidence that was knowable at the requested time.
Veil keeps deterministic astronomical state distinct from symbolic interpretation. Arcadia, the Emulator Suite, and C64 XR keep exact machine state separate from rendering, diagnosis, remastering, and agent assistance. The Agentic AI Glossary & Learning Center may receive a newer concept artifact without replacing the source-reviewed canonical entry until it passes acceptance.
These are not unrelated edge cases. They are the same architectural question expressed in different products: which output is a proposal, and which state is official?
06 / What to measure
Tool-call success is one metric inside a much larger proof.
- specialist discovery precision and capability match
- delegation and authorization enforcement
- evidence freshness, rights, scope, and citation support
- approval validity and meaningful user comprehension
- duplicate-action suppression and retry safety
- postcondition correctness, not provider status alone
- receipt completeness and state-hash agreement
- accepted-versus-latest artifact behavior
- rollback, recovery, and deterministic replay
A trustworthy evaluation campaign deliberately injects all nine failures and proves the system ends in the right state. A dashboard full of green model and tool spans is not enough when the wrong thing became official.
Working proof
Run the same pipeline against all nine failures.
The browser laboratory exposes successful A2A and MCP steps, then shows the first authority boundary that blocks, holds, deduplicates, quarantines, or rolls back the action.
Open the governed interoperability lab ↗