I specify behaviour as actions and transitions, not only tone. That includes when an agent acts, asks, escalates, refuses or defers; how uncertainty changes capability; and how past turns become behavioural precedent.

I treat context as a scarce behavioural budget. Stable instructions, contextual state and volatile turn data have different lifetimes and caching properties.

I design evals as measurement systems, not score dashboards: dataset contracts, isolation level, rubrics, verdict rules, repeated runs and calibration.

I design agent systems as controlled compositions of probabilistic and deterministic operations, each with a clear accountable boundary.

I work with retrieval as a multi-stage decision system. Query intent determines the path; heterogeneous recall sources contribute candidates; rank fusion avoids comparing incompatible score scales; reranking adds pairwise relevance; business constraints operate only inside explicit semantic bounds.

I separate prompt guidance from machine-enforced control. Safety spans input classification, route topology, tool authorisation, output verification, credential isolation, data minimisation and recovery behaviour. A component-level pass is not enough when an exception path can downgrade the system verdict.

I do not treat Anthropic Messages, OpenAI Responses, Chat Completions and cloud inference surfaces as interchangeable transports. They differ in tool schemas, structured output, caching, reasoning-state continuity, error shape, token accounting and retry behaviour. Those differences belong in a capability layer, not scattered through callers.

I specify observability around decisions, not only requests. A useful trace identifies the node, model, prompt version, tools, latency, tokens, cache behaviour, safety route and outcome while redacting the content a team does not need. Economics should follow completed tasks, including retries and rework, rather than flatter the system with cost per call.