Series
Agentic AI: How AI Agents Work
How AI agents work, measured on the Claude API: the agent loop, tool calling, planning, context engineering, MCP, multi-agent systems, coding and computer-use agents, evals, reliability and cost. Read them in order, or jump to whatever you're curious about.
What an AI Agent Is: The Loop, Measured
An agent is a model in a loop that picks tools and decides when it's done. Build one in 160 lines, trace a run, and measure tokens, cost and latency.
Agents & Tool UseHow Tool Calling Works Inside the Model
Tool schemas become prompt tokens, and a tool call is a predicted token. Strict mode, parallel calls and tool descriptions, measured on Claude Haiku 5.5.
Agents & Tool UseReasoning and Acting: ReAct, Planning and Thinking, Measured
ReAct, plan-and-execute and adaptive thinking on one 36-task tool set: accuracy, tokens and cost on Claude Haiku 5.5, and where more reasoning stops paying.
Agents & Tool UseReflection and Self-Correction in Agents, Measured
Self-critique, test feedback and Reflexion on 40 coding tasks with Claude Haiku 5.5: what each one fixes, what it breaks, and what a fix costs.
Agents & Tool UseContext Engineering: Keeping a Long Agent Run Inside the Window
What fills an agent's context window, and six ways to manage it, measured on 22 long tasks with Claude Haiku 5.5: accuracy, peak tokens and cost.
Agents & Tool UseMCP: Connecting Agents to Tools
How the Model Context Protocol works on the wire, how to plug MCP servers into an agent loop, and what happens to tool choice as tools pile up.
Agents & Tool UseMulti-Agent Systems: Orchestrators, Sub-agents and A2A, Measured
One agent or a lead with sub-agents? Measured on 26 tasks with Claude Haiku 5.5: accuracy, token multiples, cost and wall time, plus A2A on the wire.
Agents & Tool UseCoding Agents: How They Fix Real Bugs, Measured
How coding agents solve SWE-bench-style issues with search, edit and test tools, measured on 28 bugs from real Python repos with Claude Haiku 5.5.
Agents & Tool UseComputer-Use and Browser Agents: Pixels Versus the Page
How Claude's computer and browser toolsets work, what a screenshot costs in tokens, and 20 browser tasks run three ways on a sandboxed intranet.
Agents & Tool UseResearch Agents: Web Search, Fetch and Citations, Measured
Claude's web search and fetch tools on 30 fresh questions: one search vs a research loop, cost per answer, and a hand audit of 60 citations.
Agents & Tool UseEvaluating Agents: pass@k, pass^k and Run-to-Run Noise, Measured
SWE-bench, τ-bench, OSWorld and GAIA explained, then 1,140 repeated agent runs: pass@k vs pass^k, run-to-run spread, cost per solved task and judge agreement.
Agents & Tool UseLong-Horizon Reliability: Compounding Errors, Retries and Checkpoints
Why long agent runs fail more often than their steps do, measured on Claude Haiku 5.5, and what retries, idempotency keys, checkpoints and budgets fix.
Agents & Tool UseSandboxing and Permissions for AI Agents, Measured
What a sandbox costs per tool call, what it keeps out, and how many tool calls each allow/ask/deny policy interrupts, measured on 100 agent runs.
Agents & Tool UseTraining Models to Be Agents: Trajectories, Environments, Rewards
How models learn to act: fine-tuning on agent trajectories, sandboxes by the million, rewards a program can check, multi-turn RL and reward hacking.
Agents & Tool UseCost and Latency of Agents: Caching, Effort, Models and Batch, Measured
Cost per completed task and time per task on one 36-task set, lever by lever: prompt caching, effort, Haiku vs Sonnet, a Sonnet lead and the Batch API.
Agents & Tool UseAn Agent End to End: A Docs-Maintenance Agent, Measured
One docs-maintenance agent built from everything this series measured, run 612 times on real docs with seeded defects: what it finds, misses and costs.
Agents & Tool Use