Overview — Project 5: Forge

The earlier projects used models inside workflows. Forge is the first agent: the model can choose among tools, observe results, repair output, and decide what to do next. That freedom requires stronger boundaries.

You will also study the maintained OpenAI Codex CLI, which is implemented in Rust, as a real-world example of an AI agent that runs locally and coordinates configuration, terminal UI, tools, approvals, sandboxing, and model interaction. The exercise uses OpenAI’s public source as evidence; it does not ask you to imitate undocumented internals.

Run Forge

cd rust/rust-ai-engineering
cargo run -p forge-cli -- run \
  --task fixtures/forge/task-valid.json \
  --responses fixtures/forge/responses-valid.json \
  --trace target/forge-valid.jsonl

cargo run -p forge-cli -- replay --trace target/forge-valid.jsonl
cargo run -p forge-cli -- eval --suite fixtures/forge/eval-suite.json

The six days

Day	Build	Rust concept	Agent concept
18	source-guided Codex anatomy	workspaces, async components, native binary	agent = model plus controlled runtime
19	narrow tool registry	traits, enums, generics, effects	tool schemas and untrusted observations
20	progress-aware loop	state machines, budgets, termination	stop on evidence, not narration
21	approval and receipts	capabilities, typestate, policy	permission before side effects
22	causal trace, replay, memory	durable events, conflicts, hashes	debug and remember with evidence
23	release gate	datasets, reports, version identity	ship behavior, not a prompt file

Product rule

model proposes action
      ↓
harness validates syntax, policy, budget, and authorization
      ↓
tool executes narrowly
      ↓
receipt plus independent state becomes evidence
      ↓
loop may claim progress

Start with Day 18 — Learn from Codex CLI’s Rust Codebase →.