Memory and control
Context Engineering for AI Agents That Work Across Real Systems
A practical guide to designing durable context, retrieval boundaries, working packets, and policy-aware memory for AI agents.
Field notes from the control plane
Practical guides for operators who need durable context, visible execution, and publication gates that stay closed until a human approves the work.
For a run that needs human approval, start with the approval walkthrough, then inspect the conceptual AgentRun lifecycle. The example uses a mock executor. Public API access is not available.
For context and execution records, choose a guide below. Context engineering explains what information may enter a run. Observability explains how to trace the decisions and the outcome.
The policy archive documents historical gate decisions. Its measurements belong to the dated archive; use the September evidence snapshot for the later published record.
Memory and control
A practical guide to designing durable context, retrieval boundaries, working packets, and policy-aware memory for AI agents.
Evidence and operations
A field guide to tracing agent decisions, tool calls, policy gates, failures, and human approvals without confusing activity with reliability.