Summary: OpenAI Codex is an autonomous coding agent that introduced Goals — persistent, thread-scoped objectives that define "what done looks like" and iterate until evidence confirms completion. Goals transform Codex from a prompt-response tool into an evidence-driven autonomous agent.

Overview

Property Value
Provider OpenAI
Type Autonomous coding agent + Goals API
Key Innovation Persistent objectives with evidence-based termination
Scope Thread-scoped (not global memory)
Verification Benchmark/test-driven outcomes

Goals: The Mental Model Shift

PROMPT:  ask → work → result → wait (you drive)
GOAL:    work → check → continue or complete (evidence drives)

Prompt vs Goal Example

Approach Example
Prompt "Fix the checkout latency"
Goal Reduce p95 checkout latency below 120ms on the checkout benchmark while keeping the correctness suite green

Anatomy of a Strong Goal (6 Elements)

Element Purpose Example
1. Outcome End state p95 checkout latency < 120ms
2. Verification How to prove it checkout benchmark
3. Constraints What must hold correctness suite green
4. Boundaries Files/tools allowed checkout service, benchmark fixtures, related tests
5. Iteration Policy Next action selection record what changed, benchmark result, next experiment
6. Blocked Stop What to report when stuck attempted paths, evidence, blocker, next input needed

Research Workflow Example

Weak: /goal Reproduce Deep Hedging paper

Strong:

/goal Produce the strongest evidence-backed reproduction of Buehler et al. "Deep Hedging" using available materials. Attempt every headline result, verify outputs, and end with a report separating:
- Confirmed mechanics (exact replay)
- Approximate trained results (close numerical match)
- Blocked claims (missing seeds/checkpoints/data)
- Remaining uncertainty

Results achieved by this Goal:

  • ✅ Rebuild pricing/hedging mechanics
  • ✅ Reproduce Heston reference price
  • ✅ Train CVaR hedge policies
  • ✅ Rebuild main figures
  • ✅ Honestly label what couldn't be exactly replayed (missing seeds, TF graph, checkpoints)

Key insight: The Goal prevented overclaiming. Final audit preserves epistemic levels instead of flattening to "reproduced ✓".

Architecture Insights

  • Thread-scoped state — Goals live with context (files read, commands run, diffs, reasoning trail)
  • Continuation — Only at safe boundaries (idle thread, active Goal, within budget)
  • Budget limits → Summaries, not fake completions
  • Pause/resume — Preserves objective

When NOT to Use Goals

Avoid Goals For Use Goals For
One-line edits Debugging with evidence (flaky tests, perf)
Vague finish lines ("make it better") Multi-step refactors with test gates
Hiding uncertainty Research with clear evidence standards
Any task where next action depends on what you just learned

Related

Sources