Paying for Codex by the month? Make each workflow step a headless harness run

#What changed

Sergii Makarevych (@sermakarevich) published a build note on an open "agentic factory" that composes headless coding-agent runs — Codex CLI non-interactive (codex exec), Claude Code (claude -p), or OpenCode — into Temporal workflows. Manual commands, schedules, and event watchers all start the same graphs. Each step returns a schema-checked object the workflow can branch on. Repository: https://github.com/sermakarevich/agentic_factory

#Who is affected

People who already pay for Codex on a ChatGPT plan (or Claude Code / OpenCode Go) and want unattended multi-step research, report, or coding pipelines without putting an API key into a graph library.

#What to do now

1. Make one headless harness run the smallest step, not a raw model HTTP call. OpenAI’s ChatGPT-plan help says plan usage is spent through Codex and allowed partner apps, not through an API key; Codex’s noninteractive docs cover codex exec with --output-schema for structured results.
2. Require the agent to submit results against a JSON Schema while the session is still open, and loop until the check passes, so the orchestrator gets data instead of prose.
3. Put an orchestrator in front for retries, per-provider queues that pace five-hour usage windows, and schedule or event starts (the article uses Temporal).
4. Keep comparing models at API list prices in your ledger if you want, but expect the bill to stay the flat subscription fee when you stay inside the harness.

#What is not confirmed

The builder’s estimate that subscription tokens cost about 4–8× less under full use, and all token and dollar figures in the article, come from their October 2026 runs. Dotsfeed has not reproduced the factory. OpenAI does not publish a subscription-versus-API token price ratio.

Source: https://x.com/sermakarevich/status/2110374100791074924