Done needs
evidence.

A completion-time critic checks disposable copies of the current and initial environments. If it rejects the result, the original agent resumes work.

Across 89 TerminalBench 2.1 tasks and five trials, reported success rises from 63.4% to 72.4%, a 9-point gain with a reported confidence interval of 5.6–12.6 points. The MCTS comparison reaches 64.0%; model cost is 2.2× the baseline.

Evidence & limits. No significant benefit was established for the stronger Astra/Opus configurations. Forking an environment does not isolate external databases or services; destructive checks need additional isolation. Submitted October 14 EDT, then announced in the October 15 evening batch.

Query-relevant, reusable document embeddings are compressed into 16 latent memory tokens for a frozen language model.

The method beats the best tested baselines on five of eight benchmarks. On FactKG, it scores 84.33 versus 80.43 for full context while using 108× fewer document-representation tokens. TextWorld rises from 18.6 to 24.2.

Evidence & limits. The roughly 194-million-parameter memory module is trained, with benchmark-specific fine-tuning. The experiments use Qwen encoders and decoders; cross-family transfer and coding use are unproven. Fine details can be lost. The token reduction is not a total-cost claim.

A study of ToolNet and H-EPM tests nine open models on ToolSandbox and BFCL v4, with three runs per setting.

Reliable H-EPM gains appear in only four settings. On ToolSandbox, it averages 2.72× the language-model calls and 2.21× the tokens. Random or absent next-tool suggestions preserve much of the gain.

Evidence & limits. Lightweight textual situational insights perform better in most settings with almost no extra inference calls, but offline construction still costs compute. The critique concerns tool nodes and transition edges, not every entity graph or state graph.

SPACE exposes Firecracker microVM infrastructure for agent execution, including lifecycle management, pause, resume and recovery.

The Apache 2.0 release separates network control and credentials from the untrusted sandbox. The implementation is inspectable infrastructure for long-running agent workloads.

Evidence & limits. Perplexity reports zero host escapes across 108 attempts, but initial network bypasses were found and fixed. A vendor-reported test is not a guarantee of isolation across deployments or future attacks.

Drives provide persistent files independent of a Vercel Sandbox, with forks for parallel agents. General availability covers Hobby, Pro and Enterprise plans.

A Drive supports one writable mount. Read-only mounts are point-in-time snapshots, and the Drive and Sandbox must be in the same region.

Evidence & limits. Forks do not merge back automatically. Persistent storage solves workspace lifetime; coordination, conflict resolution and integration remain application responsibilities.

Parallel coding agents work from hypotheses and evaluation rubrics. They share successes and failures at cycle boundaries, reset context during handoffs, and produce ablations and claim-to-evidence mappings.

The system keeps human questions and validity judgments in the loop. The release illustrates how structured experiment cycles can preserve useful findings across parallel work.

Evidence & limits. Author-reported research includes observed reward hacking. The scope is computational machine learning; literature search supports writing rather than ideation. This is an October 15 release with an older October 13 paper, not a new paper that day.