Deliberate-then-Act: self-critique loops for high-stakes decisions
A framework where agents draft, critique, and revise a plan before committing — cutting downstream errors by 38% on operational benchmarks.
Read paper
Research
Pushing reasoning
Pushing reasoning
to the edge.
Our research lab studies how agents deliberate, use tools reliably, and stay governable as they take on real authority. Here's what we're working on and what we've published.
0
peer-reviewed papers
0%
avg. reasoning-error reduction
0
open benchmarks released
0
median edge inference
Focus areas
Four lines of inquiry
Reliable function-calling under partial failure
How agents recover when an API times out, returns malformed data, or partially succeeds — with verification and graceful rollback.
Read paperSub-second reasoning on live operational streams
Distillation and caching techniques that bring multi-step reasoning to 120ms median latency at the point of decision.
Read paperConfidence-gated autonomy
Calibrating when an agent should act, ask, or escalate — so authority scales with demonstrated reliability.
Read paperExplanations as a first-class output
Training agents to produce human-auditable rationales that hold up under scrutiny — without degrading task accuracy.
Read paperSparkBench: an open benchmark for workflow agents
200 realistic enterprise tasks spanning CRM, ticketing, and analytics — open-sourced for the community to build on.
View benchmarkRead with us
Get new papers, benchmarks, and lab notes in your inbox. No noise — just the research.
Subscribe to the lab