Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents

TL;DR

Independent runs of the same autoresearch agent on the same long-horizon task plateau at substantially different scores — gaps that persist even with more compute — because trajectories get stuck in localized regions of solution space the authors call “idea basins.” Fork-and-flush is a simple periodic intervention: at a fixed interval, fork the agent into parallel trajectories that each inherit the accumulated workspace but start with a fresh chat context, run each for a fixed horizon, then continue from the highest-scoring one. Across 13 long-horizon research and engineering tasks (individual runs lasting up to several days), fork-and-flush beats single-run by 66.0% and best-of-N by 44.4% in relative min-max-normalized average score at equal compute budget.

Key claims

Independent runs of the same autoresearch agent on the same task plateau at substantially different final scores, and the per-run gap does not close with additional compute [§Abstract].
Embedding candidate artifacts by functional similarity shows trajectories remain in localized regions of solution space (“idea basins”), providing mechanistic evidence that persistent score gaps come from basin lock-in rather than variance [§Abstract].
Fork-and-flush periodically forks the agent into parallel trajectories that inherit the accumulated workspace (filesystem, artifacts, logs) but start with a fresh chat context; each trajectory runs for a fixed horizon and the agent continues from the highest-scoring one [§Abstract].
At equal compute budget across 13 long-horizon research and engineering tasks (runs up to several days), fork-and-flush beats a single-run baseline by 66.0% relative on min-max-normalized average score [§Abstract].
At equal compute budget, fork-and-flush also beats a best-of-N baseline (running N parallel trajectories to completion and picking the best) by 44.4% relative on the same metric [§Abstract].

Method

Fork-and-flush treats a single long-horizon autoresearch run as a sequence of fixed-horizon phases rather than one monolithic context-growing trajectory. At each phase boundary, the agent is forked into several parallel copies. Each copy inherits the accumulated workspace — the full filesystem, code, experimental logs, and artifact history — but starts with a fresh chat context, discarding whatever conversational scaffolding had been building up. Each fork runs for a fixed horizon of compute and the orchestrator continues from the highest-scoring trajectory, repeating the fork-and-flush step at the next boundary.

The intervention is deliberately simple: no learned proposer, no explicit diversity objective, no cross-trajectory communication. The claim is that the chat-context reset alone is sufficient to escape idea basins because the shared workspace carries the useful state (what’s been tried, what works) while the fresh context removes the conversational commitments and framings that trap the agent in a local region of solution space.

Results

Evaluated across 13 long-horizon research and engineering tasks, where individual agent runs last up to several days. On min-max-normalized average score at equal compute budget, fork-and-flush delivers a 66.0% relative improvement over a single-run baseline and a 44.4% relative improvement over best-of-N (running N parallel independent trajectories to completion and picking the best) [§Abstract].

Why it’s interesting

Sits squarely in the AI-for-AI Research cluster but attacks a different axis than the prior filed systems: FARS, ASI-ARCH, and Karpathy autoresearch all specify the scaffold shape (role-split, shared memory, pre-experiment gate), while this paper specifies the scaffold schedule — when to interrupt and partially reset a long-horizon trajectory. Directly engages the concept page’s open question about whether parallelism changes what search the agent does, not just how fast: the embedding-space idea-basin finding is the first filed mechanistic evidence that single-run autoresearch is bottlenecked by trajectory lock-in rather than compute. Contrasts with the composition pattern in optimize_anything Goes omni — Composing Optimizers into Meta-Optimizer Pipelines — which runs different optimizers in parallel under 1/N budget — by instead running copies of the same agent in parallel with only chat-context reset as the diversifying lever, while sharing the workspace rather than resetting it. The 44.4% gain over best-of-N under matched compute is a precise cost of running trajectories to completion rather than cheaply branching at plateaus.