Building agents

Most readers first extend the agent they already use: they add a tool, write a Skill or connect an MCP server. Writing your own agent adds the loop, safety and long runs; Codex and Pi come last as complete worked examples. Splitting work across several agents is covered in the coding group, in Multi-Agent Coordination.

Reading Order

Step	Article	Your job
1	Harness: Controlling an Agent’s Execution	See where execution control, tools, state and verification sit, and where a new capability plugs in.
2	Tool Contracts for AI Agents	Add a tool: define its inputs, authority, side effects and what it returns on failure.
3	Bounded Agent Loops	Write your own loop: step limits, tool validation, result checks and cancellation.
4	Agent Security: Prompt Injection, Permissions, Sandboxes, and Credentials	Assume prompt injection sometimes succeeds, and limit its consequences with permissions, approvals, sandboxes and credential isolation.
5	Agent Skills	Package a repeatable procedure as a Skill, and check that it triggers when it should and does the right thing.
6	Model Context Protocol (MCP)	Connect or write an MCP server: architecture, lifecycle, transports and trust boundaries.
7	Long-Running Agents: Checkpoints, Recovery, and Idempotency	Let a task resume after an interruption without repeating side effects on retry.
8	Codex Harness Architecture: State, Compaction, Interfaces, and Security Boundaries	Case: how Codex handles runtime state, context compaction, integration surfaces and two layers of security control.
9	Pi Harness Architecture and Extension Boundaries	Case: Pi’s minimal core, customization resources, sessions and extension boundaries.

My reading