Coordinate remote specialists without losing accountability

This worked application uses explicit planning assumptions to demonstrate FIELD. Quantities below are inputs, calculated implications or proposed targets; they are not reported customer results. The engagement decision remains conditional on the evidence described.

Planning basis and engagement boundary

F — Frame the outcome

Use independent specialists only when they improve accepted analysis enough to justify partner latency, cost and data exposure.

Value measure: Accepted analysis cost/time and coordinator/partner exception workload.

I — Investigate the workflow

Establish peer identity, task/result contracts, clarification handling and duplicate-submission semantics. Determine what happens when one specialist is unavailable.

First evidence task: Lose the acceptance response and return a completed artifact with an incorrect total.

E — Engineer the solution

Use A2A mappings, ADK persistence audit, artifact verification and observation traffic budgets; keep final business authority in the coordinator.

The implementation assessment below makes the authority, workload, resource and failure assumptions concrete. Its numbers are calculations from the stated scenario, not a vendor performance guarantee.

Implementation plan

Keep coordination in your harness and use A2A where the peer owns a remote task lifecycle. Use MCP inside either agent for tools where useful; do not expose internal framework objects as your interoperability contract. Each delegation carries a bounded objective, authorized input references, output schema, deadline and permitted data class. A peer registry holds approved identities, contract versions and routing rules. A public card can discover a candidate, but the registry/policy determines whether customer data may be sent.

Maintain local delegation records mapping run/step to peer context/task/message/artifact identities. Store outbound intent before submission. For acceptance timeouts, use the peer's verified deduplication/lookup contract. If absent, represent uncertainty rather than launching another write-capable job. The coordinator joins independent results using explicit completeness and conflict rules. A policy interpretation and a research result are evidence inputs; the final business action still needs local verification and authority.

For ADK, inspect session, artifact, memory and A2A task/push stores individually. Replace volatile defaults for the required failure model. A durable session service does not by itself guarantee cross-replica A2A task recovery. Pin SDK and wire versions; compare the actual generated card/schema against the peer profile. The framework-specific source is the ADK A2A wrapper reference.

Resources and waiting

Two incoming tasks/s × two delegations = four remote jobs/s. With a 90-second mean lifetime, expect 360 active remote jobs in a stable system. Polling every ten seconds produces roughly 36 status requests/s before retries and submission traffic. Long tails and twenty-minute jobs increase active state; size from the duration distribution rather than only the mean. A shared scheduler can poll without one thread/job. Streaming/push choices need capability checks, connection budgets and outage reconciliation.

If each job returns a 2 MB artifact, steady returned payload is about 8 MB/s before replication or user downloads. Prefer content-addressed object references with authorized retrieval, not giant JSON messages copied through every coordinator event. The coordinator's CPU budget depends on validation, joins and parsing, while remote model computation is a separate cost/quota boundary.

Failure and evaluation

Persist input-required state and release compute while waiting. On a user reply, verify the correct task and unchanged authority. If one peer times out, decide whether the customer contract permits a partial answer; never let a stale success event overwrite a later authoritative terminal outcome through an unversioned local merge. Reconcile after a notification outage. Bound recursive delegation and reject requests for broader data than originally authorized.

Test missing artifacts, conflicting conclusions, malformed schemas, lost acceptance responses, duplicate notifications, restart during clarification and peer version upgrades. Score verified answer completeness, citation quality, deadline attainment and cost per accepted result. A second model can help assess semantics but cannot establish protocol or authorization correctness.

If the scope changes: Both specialists are owned by the same team, finish in 100 ms and deploy together. Local function or in-process composition may remove unnecessary network, discovery and lifecycle complexity. Preserve interface boundaries in code so remote separation remains possible later.

Simpler alternative: A durable deterministic workflow invokes two ordinary service APIs if the remote systems do not need agent-specific discovery or conversational task continuation. Explain why A2A adds value before mandating it.

L — Launch with evidence

Exercise lost acceptance, missed notifications, restart during clarification and wrong artifacts; compare with a simpler local/service workflow.

Critical decision condition: Protocol completion alone cannot establish artifact correctness or permission for consequential action.

Attach raw artifacts, the tested configuration, population, denominator, date and limitations to the release record. Missing or stale evidence holds the dependent decision. A favorable value forecast cannot offset a failed critical condition.

D — Develop ownership and reuse

The coordinator owner manages partner contracts and ADK operators own persistent state; define who resolves ambiguous remote submissions.

Candidate reusable asset: Remote delegation contract and artifact verifier.

Commercial decision and next commitment

Accountable decision owner: Workflow owner with coordinator/partner service owners.

The initial commitment is the bounded journey above, with the stated exclusions. Confirm access, customer staffing and acceptance ownership during discovery; price and schedule unresolved dependencies explicitly. A new source, tenant class, language, action or operating envelope requires an impact review rather than silent scope growth.

Approve only the next exposure level supported by the evidence. The next useful investment is: lose the acceptance response and return a completed artifact with an incorrect total.

Method and references

Apply the FIELD method, economic model and evidence rules. Methodological and technical influences are listed in the source register. The protocol assessment follows the agent control chapter; pin and verify the actual protocol/runtime combination before implementation.