Codex users report conflicting results on long-running autonomous tasks

• Astra and Sol repeatedly stop at plans or promises instead of continuing multi-step work, according to BEBischof's account.
• Six-hour autonomous runs work for svonava, using adjusted GitHub permission rules and skills for verification and compute management.
• The /goal workaround can chase the wrong objective on broad projects and waste tokens, BEBischof says.

OpenAI's own long-horizon experiment ran GPT-5.3-Codex for roughly 25 hours building a design tool, with verification after each milestone. Codex already uses Lightroom's scripting capabilities, according to BEBischof, illustrating the desktop-app work involved in this discussion.

BEBischof describes a regression from earlier Codex sessions with Sol, when large projects kept moving across multiple steps. With Astra at any thinking level, and with the version they call “6.1 sol,” the recurring failure is announcing the next action without taking it.

svonava reports six-hour stretches without permission requests after adjusting GitHub permission rules, supported by verification and compute-management skills. Their inference-system work, including LoRAs, uses several orchestration patterns:

• Nested CLIs: An outer, long-running Codex CLI manager coordinates a hypothesis DAG and scheduler, with separate inner Codex executions.
• Built-in subagents: More recent runs use less explicit structure and delegate through Codex's subagents.
• Sibling CLIs: Two or three Codex instances coordinate through shared read/write state, depending on verification costs and bottlenecks.

svonava still discovers inefficient or incorrect work on returning; the reported progress includes improving a metric or testing hypotheses.

BEBischof's projects include work with a different shape:

• Process a million files, then operate desktop applications.
• Run mathematical searches over data.
• Run simulations and vary parameters based on the results.

Those jobs could be reframed as software with a verifier or metric, svonava suggested. The expensive catch, BEBischof argued, is that the reframing could add “maybe 100x complexity” and consume scarce attention building tools rather than solving the original problem.

The /goal workaround helps with strongly scoped tasks, according to BEBischof. On broad, multi-part projects, they say it can fixate on the wrong objective and chase token-consuming hacks.

OpenAI's Goals cookbook describes concrete continuation rules:

• Thread-scoped state: The objective, lifecycle and budget persist within the thread.
• Idle-thread dispatch: Continuation requires an active goal, remaining budget, and no queued user input or pending thread work.
• Plan-only stops: Plan-only work does not trigger continuation.
• Tool-call guard: If a continuation turn makes no tool call, the next automatic continuation is suppressed.

BEBischof's migration request includes preserving context and good mobile support. OpenAI's remote-work guide describes the phone as a control interface while execution stays on a connected host.

kevinkern estimates that 80% of agent traces are useless to them and aggregates “needs you” items across machines.

In a separate complaint, BEBischof said Booking.com's AI phone support could do little and made reaching a human almost impossible.

BEBischof also questioned AI helpline resolution figures in a follow-up, arguing that refused escalations and hang-ups could inflate apparent success.