What it covers
- Planning and orchestration — Multi-step plans, tool selection and hand-offs between specialised agents, with a deterministic controller deciding what runs next.
- Context, memory and retrieval — Retrieval-augmented generation over your documents with citations, per-tenant memory, and refusal when nothing relevant is found.
- Tool calling and human approval — Typed tool contracts, permission scopes per action, and approval gates where money, data or customers are affected.
- Durable execution and recovery — Long-running jobs that survive restarts, idempotent steps, retries with back-off and a dead-letter path someone actually reads.
- Evals and observability — Fixture transcripts with known triggers, pass/fail logs per prompt version, and cost per operation from the first week.
- Security and deployment — Prompts assembled server-side so they can't be tampered with; keys, logs and inference in your cloud, in the region your data must stay.
How we work
Deterministic code does the deterministic work — classification, scoring, retrieval, routing — and the model does interpretation and language. Prompts live in version control with tests. We report cost per operation and containment rate from the first week, and every agent action is logged so a person can audit what it did and why.
