RSCH
Research
Long-form investigations into how AI agents behave in real engineering work. Each programme is evidence-labelled, sequenced as a learning path, and ends with something you can do on Monday.
What every research programme gives you
- A one-page summary to decide in five minutes whether it matters to you.
- A learning path in parts, from the problem to a plan you can run on Monday.
- Evidence labels on non-obvious numeric or contestable claims; a sources page that lists verification limits.
- Diagrams, charts, templates and code you can copy.
- An FAQ and a glossary for the questions and terms that come up most.
Programmes
RSCH-001PublishedContext Management in Coding AgentsChroma's controlled retrieval and question-answer tests found performance generally fell as input grew, with model- and task-specific curves. Agent studies compare specific compaction strategies. This programme is an evidence synthesis; it did not test a direct coding-agent window-size effect or establish a universal window-size rule.Open the research →
In preparationLoop EngineeringHow is an agent's iterate-and-repair loop built, and what makes it converge instead of spin?
In preparationAgent EvaluationHow do you know an agent actually got better, with what design, at what cost?
In preparationPrompt Cache ArchitectureWhat does editing a file mid-session really cost once the prompt cache is priced in?
In preparationMulti-Agent Maturity ModelWhen is a team actually ready for multi-agent systems, and what has to be true first?