A2A ARCHITECTURE · FAILURE ISOLATION
A2A architecture: isolate one child-agent failure
Fan one request out to three remote agents, break the stock agent, and see whether the join keeps or throws away the healthy results.
Event trace
Computed data
Final metrics
Teaching model of an orchestrator delegating three tasks over A2A; no network, agent card or JSON-RPC call is made. Child durations are fixed (weather 1.0 s, catalog 2.5 s, stock 2.0 s, backup 1.5 s), a retry is a new task sent immediately, and a reply exactly at the timeout counts as on time. Streaming updates, partial artifacts inside one task and cancellation latency are not modeled.
One failed child, three lost results
The page opens with Stock agent fault set to First attempt fails, Retries per child at 0 and Join policy All-or-nothing. Weather completes at 1.0 s, stock fails at 1.0 s, and the orchestrator fails the whole cart request at 1.0 s. Catalog keeps working and completes at 2.5 s for nobody: REQUEST FAILED, 0/3 artifacts returned, 2 healthy discarded, and 1.5 s of orphaned work in the metrics.
Isolate the failure instead
Choose Return partial results: weather and catalog come back at 2.5 s and stock is reported as failed, ISOLATED with 2/3 artifacts. Fail fast and cancel siblings still fails at 1.0 s, but it cancels catalog instead of orphaning it, so the metrics list catalog under canceled and show 0 s of orphaned work. With Retries per child at 1, stock-2 succeeds at 3.0 s and every policy returns COMPLETE, 3/3.
Timeouts and the deadline
Pick Never replies: the Child task timeout of 4 s cancels stock, and partial results return 2/3 at 4.0 s. Lower the timeout to 2 s and catalog, which needs 2.5 s, is cut too: only weather survives, 1/3 at 2.0 s. A timeout tighter than a healthy child is a failure of its own. Replies after 7 s with a 10 s timeout completes at 7.0 s; lower Orchestrator deadline to 5 s and stock is cancelled unfinished at 5.0 s.
A backup agent
Pick Every attempt fails with 2 retries: stock fails three times and all-or-nothing fails the request at 3.0 s. Tick Backup stock agent after the last retry and the backup answers at 4.5 s: COMPLETE, with 4 attempts in the stock row. The backup needs 1.5 s, so with a 1 s timeout it times out as well, and catalog fails on every attempt.
Replay and limits
Any change recomputes the run and rewinds it; the Complete-run preview states the outcome at once. Play and Pause walk through the events, Next event jumps one event, Restart rewinds, Speed runs from 0.25x to 2x and Inspection time seeks from 0 to 60 seconds, where the outcome and final metrics appear. Child durations are fixed (weather 1.0 s, catalog 2.5 s, stock 2.0 s, backup 1.5 s), a retry is a new task sent at once, and a reply exactly at the timeout counts as on time. No agent card, JSON-RPC call or network is involved.