Flat agent team: direct handoff or broadcast

Task trace

What this lab demonstrates

A flat organization has no manager in the path of every decision. Each specialist can accept a request directly from a requester, and peers may pass work among themselves. This model has four explicit skills: research, verification, writing, and release approval. Each task names the peer that owns it. Direct handoff sends one message to that peer. Broadcast sends a copy to every peer, even though only one has the matching skill. The message count therefore rises without inventing extra completed work. This is a communication comparison, not a claim about model intelligence.

How to use each control

Peer coordination chooses direct handoff or broadcast. Unavailable peer removes one named specialist from the fixture; choose No unavailable peer for the baseline. Play sequence moves one task through the visual timeline every fifteen seconds. Pause sequence freezes the current task, Restart goes back to zero, and the Timeline slider jumps to a particular point. Speed changes viewing time, not routing results. Read each card's route and exact outcome, then use the text trace below the simulation to compare all four tasks even if motion is paused.

Try a controlled comparison

Start with no unavailable peer and direct handoff. All four tasks should complete with four messages. Switch to broadcast; the same four tasks complete, but sixteen messages are sent because four peers receive each of four requests. Now remove the verification peer. Verification becomes blocked under either coordination setting: broadcasting to more peers does not manufacture the missing skill. Pause at thirty seconds to inspect the first two tasks and compare the visible route against the final trace. What would a real team need to add for substitution or escalation?

What the model simplifies

The example uses four fixed tasks, one skill per peer, and no queue, retries, deadlines, model output, or changing worker capacity. A message is a teaching unit, not a measured network packet or token bill. A real flat team may have overlapping skills and negotiate ownership dynamically; this fixture deliberately does not. ADK Java is not launched, and no production agent is contacted. Completion means only that the selected peer exists in this synthetic map. It says nothing about answer quality or whether approval was justified.

Inspect ownership without a coordinator hierarchy

Four task requests represent retrieving evidence, verifying a claim, writing an answer and approving release. Each has one explicitly skilled peer. Direct peer handoff sends one request to the assigned peer; broadcast sends each request to all four peers, but only the skilled peer performs the task. The simulation counts request messages, not acknowledgements, retries or gossip traffic. It does not measure token cost or claim that a flat organization is always cheaper than a hierarchy.

Select the coordination and failure inputs

Peer coordination chooses direct handoff or broadcast. Unavailable peer selects no failure or one of the four skills. Removing a peer blocks that skill’s task because this fixture contains no backup owner. The other tasks still receive independent outcomes: the displayed ordering is an inspection sequence, not a dependency graph that propagates a failed retrieval into later writing. This distinction matters when translating the example into a real multi-agent workflow.

Reproduce message and ownership counts

Start with no unavailable peer. Direct handoff completes all four tasks with four request messages. Broadcast completes the same four tasks with sixteen messages. Choose Verification peer unavailable: exactly one task becomes blocked and three remain done. Direct handoff still records four messages because the request to the unavailable verifier is counted. Broadcast still records sixteen. Broadcasting the same request more widely cannot manufacture a qualified replacement in the fixed membership list.

Test the missing-owner boundary

Repeat the failure for research, writing and approval. Each selection should move the blocked state to the corresponding task while preserving the total of four outcomes. A missing approval peer is not silently replaced by a writer. There are no elections, timeouts, load balancing, discovery changes or autonomous role acquisition. The counts express an explicit local dispatch rule over four fictional peers. A production system must decide whether later tasks may proceed without an earlier result and which peer can legitimately take over failed work; neither policy is inferred here.

Inspect the ordered task cards

Play starts the sixty-second inspection and becomes Pause while running. Restart rewinds without changing coordination or failure inputs. Speed changes viewing pace; Timeline seeks to a stage and pauses playback. Parameter changes recompute and rewind the fixture. The summary and text trace preview all four computed outcomes from the beginning, while the cards progressively change from waiting through routing to their final state. Compare the inspected-task counter with that full-fixture preview instead of treating a previewed completion as a task already reached by playback. No peer is contacted and viewing seconds are not execution latency.