Lab

Research with operational application

Three active research areas. No editorial feed, no paper theatre - questions we are actively instrumenting in systems.

Reliable agent operation

How tool-using agents remain recoverable under partial failure, network loss, credential expiry and human interruption - without losing capital, state or audit integrity.

Current questions

Halt semantics that never strand open work or funds
Approval UX dense enough for operators and light enough to use
Evaluation harnesses for multi-step tool tasks under realistic failure
Separation of read-only intel from write-capable execution paths

Intended application

Production agent runtimes for ops, research and market systems where a hard stop is non-optional.

Intelligence interfaces

Product and internal-tool patterns that convert model output into a decision object with ownership, review, versioning and audit - not a disposable chat transcript.

Current questions

Traceable decision records that span tools and sessions
Human-in-the-loop density versus throughput
Handoff protocols between agents and between agent and human
UI that surfaces uncertainty without drowning the operator

Intended application

Internal products and customer-facing surfaces where decisions must be explainable after the fact.

Adaptive systems and evaluation

Closing the loop between live signals, experiments and execution without uncontrolled automation or metric gaming.

Current questions

Attribution under platform and channel noise
Safe experiment pacing with spend and publish gates
Outcome metrics that resist vanity optimisation
Feedback that improves the next cycle without silent policy drift

Intended application

Growth and operations loops for AI products that must improve week over week under real constraints.

Collaboration

Lab work is primarily applied: questions are chosen because a live system needs the answer. If you have an operational environment where one of these tracks is a bottleneck, contact us.