Field notes on agentic AI and harness engineering.
How to build agents that use real apps safely: tools, context, permissions, approvals and the loop around the model.
- Workflows
Agentic workflows vs autonomous agents: which should you build?
Agentic workflows vs autonomous agents: compare determinism, flexibility, cost and failure handling, and learn when a fixed chain with LLM steps wins.
-
Harness engineering
How to evaluate AI agents that use tools (without fooling yourself)
A practical guide to evaluating tool-using AI agents: task suites, mocked vs real tools, fixtures, outcome and trajectory checks, flaky runs and release gates.
-
Security and control
OAuth for AI agents: handling tokens without leaking them
How to use OAuth safely with AI agents: authorization code with PKCE, least-privilege scopes, refresh rotation, token storage, revocation and reauth flows.
-
Agent engineering
12 ways AI agents misuse tools, and how to fix each
Twelve tool-calling failure modes, from hallucinated arguments to retry storms and prompt injection, each with its symptom and a harness-level fix.
-
Security and control
Human approval for AI agents: design it so people use it
How to design human approval for AI agents: risk tiers, previews of the exact action, expiry, safe retries, fewer alerts, an audit trail and undo.
-
Agent engineering
Context engineering for agents: what to keep in the window
Context engineering for AI agents: select what goes in the window, compact history, trim tool results, use sub-agents and order prompts so caching works.
-
Agent engineering
How to design tool schemas that AI agents use correctly
Design tool schemas for AI agents: naming, descriptions as prompts, enums, idempotency and helpful errors, with a before and after JSON schema example.
-
Agent engineering
MCP explained: how models call tools, and where it breaks
MCP (Model Context Protocol) explained for builders: how clients and servers exchange tools, transports, auth, session URL risks, context bloat and versioning.
-
Harness engineering
Harness engineering: the 7 layers around an LLM agent
Harness engineering covers seven layers around an LLM agent: tools, context, memory, permissions, the loop, evaluation and observability. What each one does.
-
Harness engineering
What is an agent harness? The part of an AI agent that isn't the model
An agent harness is everything around the model: tools, context, memory, permissions and the loop. Learn what it does and why it decides if an agent works.