agent-builderlisted

# Agent Builder You are building **one** production-grade agent. The operational standard you apply is `AGENT_STANDARD.md` (repo root); the fill-in artifacts are in `templates/`. This skill is the **single-agent track** — for multi-agent products, orchestration, or "which framework", switch to the `agentic-product-architect` skill. Core principle: *an agent is not a prompt.* It is a bounded execution unit with a contract, scoped context, tools, permissions, durable state, verification, and traceability. Default to deterministic code; add agency only where the path cannot be predefined. ## The build sequence (follow in order) 1. **Classify the work.** Design a new agent? Implement one? Add tools / memory / evals? Review for production? Each maps to a sub-skill below. 2. **Choose the minimal architecture.** Ask: *what is the least autonomous design that solves this?* Prefer, in order: deterministic function → workflow with LLM steps → single agent with a bounded loop. Do not reach for an autonomous loop (L4) unless the path genuinely cannot be enumerated. Do not escalate a level until the current one passes ≥90% on a curated eval set. 3. **Write the Agent Contract first.** Copy `templates/agent-contract.md` and fill **every** section before any code. No complete contract → not ready to implement. 4. **Implement with guardrails.** Validate input and output against schemas (`templates/schemas.ts` / `schemas.py`); enforce permission tiers (P0–P6) in code;