Labs explores the tools and methods behind our work: agentic development, knowledge retrieval, context management and evaluation.
We aim to share useful methods and findings as the work develops. Nothing here is released yet, and there is no download to collect.
Most agent failures are decided before the agent runs, in how the task was described. We are testing what a task brief needs to contain — inputs, boundaries, completion criteria, escalation — for the result to be reviewable.
An answer is only useful if you can see where it came from and know the reader was allowed to see it. We are working on retrieval that keeps derived content linked to approved sources while respecting existing permissions.
Status vocabulary: Concept · Prototype · Internal use · Released. A released item carries a usable link. A field note carries a date. We will not list either before it exists.
The runtime around an agent: state, tools, permissions, execution limits and evaluation.
Managing instructions, retrieved information, tool outputs and task state — beyond prompt writing.
Real questions and real tasks, scored consistently enough to notice a regression.
Self-hosted and hybrid setups measured on quality, cost, latency and data handling.
We are interested in comparing notes, and in the cases where these methods failed for you.
Discuss your project