A lot of people are asking how I pulled off these super long-horizon builds with Astra.
Summary
Matt Shumer uses a "Manager Loop" to enable Astra to complete difficult long-horizon tasks by splitting work into phases managed by a separate agent.
The Manager Loop automates steering.
A manager agent creates a massive checklist and phases, then spawns a Codex implementer agent in a separate thread to execute them autonomously.
the model is doing the steering for me
Astra asymptotes on long tasks.
Without intervention, Astra's progress stalls as it gets stuck in minutiae and focuses too much on small details.
Astra tends to asymptote
Wording prevents minutiae traps.
Shumer found that instructing the implementer to work "extremely well" rather than "perfectly" allows the model to move on once a phase is sufficient.
asking for "perfect" sent the model right back into the minutiae
Visual tracking forces progress.
Having the implementer update an HTML checklist with a progress chart helps the model notice when it has stopped making headway.
it forces the model to notice things like "I haven't made progress in a while
Context growth limits longevity.
Shumer expects asymptoting to return because the implementer's context grows across phases; he proposes spinning up fresh agents per phase to fix this.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.