Agentic Coding

Coding agents multiply output — and chaos, if undirected. The skill is making them reliable: clear specs, tight feedback loops, and deterministic guardrails so you're not just hoping the model behaves.

Claude Code building blocks (from the guide)

• Hooks (lifecycle: PreToolUse, PostToolUse, etc.): intercept, validate, format, and log agent actions. Use them as deterministic guardrails — block edits to protected paths, auto-run formatters, run tests after every edit, deny secrets. Don't rely on the model's goodwill; wire the policy.
• Sub-agents: delegate bounded tasks to focused agents; keep each one's context narrow; orchestrate planner -> worker -> reviewer patterns.
• MCP tools: extend the agent with external tools/data via Model Context Protocol servers — prefer typed, narrowly-scoped tools over broad ones.
• Events & payloads: know what each hook/event carries so your automation is precise, not guesswork.

Disciplines for reliable agentic coding

• Spec first, small vertical slices. Review diffs — never rubber-stamp.
• Tight feedback in the loop: tests + linters wired into hooks so the agent gets automatic signal.
• Context hygiene: give the agent only what it needs. Pull files into a workspace (the town-github-coding-workflow VFS pattern) rather than dumping the whole repo.
• Guardrails: protected paths, no secrets, no Docker (house rule), staging-first.
• Human-in-the-loop on irreversible actions; autonomous on safe, reversible ones.

His multi-agent patterns

LocalForge (Planner -> Writer -> Reviewer -> Tester), reliable-engineering-agent (disciplined agent), guardian-agent (Guardian Mesh security), autocoder/argent (local GGUF + WASM sandbox for untrusted execution). The disciplines above are exactly what make these dependable.

Tie-ins

senior-fullstack-engineer (the craft the agent must uphold) · town-github-coding-workflow (VFS -> branch -> atomic commit -> PR) · phased-delivery (plan before the agent builds) · security-zero-knowledge (sandbox untrusted agent execution — the argent WASM model).

Anti-patterns to refuse

• Letting the agent run unbounded; rubber-stamping diffs; no automated tests in the loop; dumping whole repos into context; agents with broad write access and zero guardrails.