LAB NOTES
Labs
Empirical Lab Reports on agent architecture. Hypothesis, method, results, what we got wrong. Repo + data + seed published with every report.
The poison was stored, retrieved and cited every time. What changed was what it was allowed to cause.
Storage, citation and reachability are three different measurements, and only the last one is under architectural control.
Prompt hardening moved three models three different ways. The gate moved all of them to zero.
A deterministic capability check at the call site removes reachability whatever the model does, while prompt hardening changes compliance unpredictably from model to model.
The baseline got the right answer and emitted one dispatchable action in thirty runs
An outcome metric and a trajectory metric rank the same systems differently, because an answer that reads correctly to a keyword scorer can be unexecutable by anything downstream.
A second attempt was worth six points. Deciding when to take one was worth one.
The gap Lab-001 attributed to topology is carried by the retry, not by the extra agents, and the two factors interact rather than adding.
Asking a model to write a probability throws away what it already knows
Reading a model's answer as text discards graded information it demonstrably holds, and the loss is a property of the readout rather than of the model or the prompt.
The durable agent, and whether ADK 2.0's crash-resilient runtime earns its overhead
A dynamic, human-gated multi-agent graph earns its overhead on a complex incident-triage task a static router cannot decompose, and its durable runtime recovers crashed work on resume instead of re-paying for it.
Multi-agent vs router on 100 customer-support queries
A workflow router beats a 3-agent system on accuracy for sub-5-step support queries, with lower cost and lower latency.