Google for Startups AI Agents Challenge · Track 2 · Optimize

Verified memory for agent fleets.

Agent fleets are getting good at talking to each other over A2A — and bad at trusting what they say. Sentinel Mesh sits in the message path: every claim a worker emits is judged by a 3-lens adversarial Gemini panel against memory as ground truth, money/auth claims hit a deterministic tripwire with no model in the loop, and what survives is written back as verified memory — so the fleet gets cheaper and more reliable every run.

✓ Verified served from memory · ▲ Flagged suspicious, capped · ✕ Blocked vetoed in-path — all counters measured, not asserted.
Measured on the recorded two-run demo + the 30-claim referee eval
~12,000×
faster to re-serve a verified claim 0 calls / 0.3 ms vs 3 calls / 3.7 s on Run 1
0
Gemini calls to re-serve a [VERIFIED] fact memory-first intercept — zero marginal model cost
$4,250
hallucinated "auto-approve refund" vetoed deterministic tripwire · 0 model calls
100%
referee verdict accuracy × 3 runs 30-claim eval · 10 accept / 9 flag / 11 veto
100%
run-to-run consistency deterministic harness + Google ADK eval
9 → 3
fleet Gemini lens calls between runs reliability and cost improve together, no retraining

The closed loop

The referee guards memory. Memory grounds the referee.

Three moving parts, one compounding loop. A claim flagged and corrected on Run 1 is auto-vetoed on Run 2; a claim already verified never reaches a model again.

Intercept in the message path

The sentinel binds to ADK's non-experimental before_tool_callback on a real A2A fleet (to_a2a() workers, RemoteA2aAgent orchestrator). A vetoed claim's tool never fires.

Judge against memory as truth

A 3-lens adversarial Gemini panel on Vertex (robust · overreach · nevers) rules in parallel; a deterministic regex tripwire independently hard-vetoes money / auth / destructive claims before any model can rationalize them.

Remember what survives

Accepts are written back into a provenance-tracked verified ledger — its own file, so broker refresh jobs can never clobber rulings. Run N+1 reads strictly cleaner memory.


Why this isn't another memory layer

Every agent memory today is gullible. It stores whatever an agent says.

One hallucinated claim, remembered once, contaminates every downstream agent — and a fleet sharing memory over A2A shares the contamination. Sentinel Mesh is the third category: not observability (watch it fail), not per-action guardrails (block, learn nothing) — verified memory as infrastructure.

A

Advisory vs verified are not the same trust

[ADVISORY] memory (curated, unadjudicated) caps suspicion at a flag; only [VERIFIED] rows that survived the panel justify a hard veto.

Paraphrase is never contradiction

"December 31" and "31 December" agree. The rubric scores meaning, not string distance — so honest restatement is never punished as a lie.

A deterministic tripwire no model can talk past

Money / auth / destructive claims hit a regex word-boundary gate and are vetoed before a model is even asked — the one place you cannot afford a confident hallucination.

Herd Immunity: verification is portable

A fleet exports an HMAC-signed Trust Passport; a second fleet imports it and vetoes a lie it has never seen — re-serving the inherited fact at 0 calls. A control fleet without it only flags the same lie.


We verified the verifier — with Google's own eval tooling
100% verdict accuracy × 3 runs · 100% consistency

30 claims across the rubric's edges — verbatim and paraphrase accepts, advisory-contradiction and memory-silent flags, verified-contradiction and tripwire vetoes (EN + KR). Scored by a deterministic harness and Google ADK eval (EvalSet · AgentEvaluator, GREEN ×3). The eval immediately caught a real failure mode and we tightened the rubric: 96.9% → 100%.

See the full eval breakdown
per-class · precision / recall
accept  1.00 / 1.00 · n=10
flag    1.00 / 1.00 · n=9
veto   1.00 / 1.00 · n=11
tripwire exact · negatives hold