Archives

A pattern for treating agents as governed runtime assets with identity, owner, purpose, limits, and retirement state.

A pattern for blocking unsafe agent changes before they reach production.

A pattern for evaluating agent behavior through trajectories, policy checks, groundedness, and cost limits.

A pattern for constraining agent-directed code execution with workspace, command, network, and approval policies.

A pattern for making agent memory scoped, typed, validated, and isolated across users and agents.

A pattern for enforcing authorization during retrieval and carrying source provenance into generated answers.

A pattern for treating untrusted content as data and preserving the original user goal.

A pattern for sanitizing sensitive data before it crosses model, tool, display, or audit boundaries.

A pattern for bounding model calls, tool calls, retries, tokens, handoffs, and per-run spend.

A pattern for reconstructing agent decisions through structured, redacted events without storing private reasoning.

A pattern for placing human review only where agent actions need judgment, accountability, or reversibility checks.

The first pattern in our Agent Harness series: defining a deterministic policy boundary between the model's reasoning and the execution of external tools.

An introduction to the Agent Harness series: defining why the LLM is just a reasoning engine and how the harness manages memory, tools, and execution boundaries.