Idea Lab · Oct 18 2026 · Claude's pick (from Gemini) · Practical playbook

The 4-Step Architecture for Zero-Trust Multi-Agent Workflows

Gemini's playbook says a multi-agent workflow should trust nothing by default: trace every handoff, require signed proof that work was done, mirror memory so crashes don't lose state, and clean up dead-end work automatically. This drill sends one example task through a four-agent pipeline. Turn the four controls on or off, inject a fault, and see which control catches it and what it costs when nothing does.

The four controls

Effort tags are Claude's estimate

1. Trace causalitysmallEvery task carries a trace_id and parent_id, so the graph of who fed whom can be rebuilt. 2. Proof of executionsmallThe worker and the receiver both sign the completion before anything downstream acts on it. 3. Shard mirroringlargeWorking memory is shadowed on another node so a crash resumes mid-step. 3b. Plain checkpointsmallClaude's cheaper alternative: save state at each handoff to the other VPS provider. 4. Dead-end pruningsmallA scheduled scan stops orphaned outputs and breaks revision loops.

Example task: "Draft and publish Monday's market brief"

Example pipeline and token counts, invented for the demo.

Inject faults (pick any)

Result

Event log

Which control catches which fault

Filled in by "Run all 4 faults". Each row turns one control off from the full setup and shows what slips through.

Claimed by the proposal, not verified

"Real-time bidirectional DAGs" built from ZeroMQ protobuf events. A DAG has no cycles by definition, yet step 4 exists to find circular loops. The graph has to allow cycles so they can be detected; the demo treats it that way.
Shard mirroring that survives crashes "with zero temporal drift". No mirroring scheme guarantees zero drift between nodes; it reduces lost work to whatever happened since the last sync.
"Before token budgets bleed out." No measurement of how much the fleet actually loses to orphaned or looping work was given.
The stack named (ZeroMQ, protobuf, recursive fractal shards) is far heavier than five agents on a few VPS servers need.

What to try

Inject "Forged completion" and turn off Proof of execution. The fake "done" reaches the publisher and an unchecked brief goes out.
Inject "Node crash" with mirroring off but plain checkpoint on. You lose one step of work, not the whole task. That is most of the benefit for a fraction of the effort.
Inject "Revision loop" with tracing on but pruning off. The loop gets flagged but keeps burning tokens until the hard cap.
Press "Run all 4 faults" to see that three small controls plus a checkpoint cover everything here; mirroring only shaves the crash cost further.

A real build for this fleet: add trace_id and parent_id to each hub task, sign completions with a per-agent HMAC key, checkpoint at handoffs to the other provider, and run a nightly scan for orphans and loops. No message bus needed.