Control the
whole path.

SkillBrowser lets the agent identify a capability gap during reasoning, formulate a targeted query, and retrieve the procedural skill it needs. It avoids loading a large skill catalogue up front.

Across five models and six SRA-Bench subsets, with 26,262 candidate skills, it beats the same-backend static retriever in 23 of 30 settings, with one tie.

Evidence & limits. Submitted to ACL Rolling Review, not accepted. Author-reported results; useful skills sometimes go unrequested, the catalogue is static, and latency rises. Both query formulation and exposure timing change, so the study does not isolate selective loading alone. Its timing comparison excludes the baseline’s precomputed retrieval.

TIDER combines semantic retrieval with an RL-trained 1.7B temporal-window locator, then reranks the combined evidence. Dates help decide which memories become candidates, before answer generation.

With fixed answer models and equal top-20 retrieval budgets, authors report 61.52% accuracy on LifeBench versus 52.70% for the strongest baseline, and 78.96% on LoCoMo versus 75.30%.

Evidence & limits. Acceptance is corroborated by the official NeurIPS catalogue. Experiments use synthetic English histories; LifeBench’s held-out test has 408 questions from two users. Results are three-run means, not independently reproduced. Upstream date quality matters; unanswerable and adversarial questions were excluded. Training and an extra retrieval stage add cost.

GatewayBench contains 524 tasks and 11,118 tool definitions; its main study compares seven MCP gateways on 450 DTap tasks. It compares flattened tool exposure with hierarchical discovery under both benign and adversarial tasks.

Using GPT-5.4 and the OpenAI Agents SDK with no distractor tools injected, hierarchical gateways averaged 30.9% indirect-attack success versus 47.2% for flattened gateways. Benign task success was 85.1% versus 87.2%, respectively.

Evidence & limits. Author-reported preprint results, not production breach rates. Main experiments use one run per setting, default gateway configurations and two model tiers. Distractor tools have authentic schemas but mocked execution. Effects depend on the model and harness; no independently verified code release was located.

Monitor-Validity Failures reports 18 reproduced before-and-after failures across 11 project groups. In one NeMo case, a detector recognized prohibited content, but its positive result did not reach the blocking decision.

Fourteen cases reached enforcement boundaries: seven native actions, six configured boundaries and one synthetic consequence. The research examines how changing schemas, streams and adapters can break the path that carries safety evidence.

Evidence & limits. Targeted historical repair cases, not a prevalence estimate or evidence that current releases remain vulnerable. This is a preprint and the reproductions are the author’s. Supplementary artifacts are referenced, but a linked runnable repository was not located during this review.

Letta Code’s new CodeMode combines built-in and MCP calls in short JavaScript scripts without another model call or subagent. Scripts can yield and return completion to the originating conversation.

The same release hardens compaction: a summary is committed only when message count or estimated tokens decrease. Sliding-window compaction that makes no progress falls back to full compaction, then an explicit irreducible-context error. Failed resume-state reads also stop masquerading as an absence of pending approvals.

Evidence & limits. Maintainer-reported implementation and regression tests, with no independent speed or token-savings benchmark. CodeMode uses trusted node:vm execution, explicitly not a security sandbox. Nested calls requiring approval fail closed and must be invoked directly.

Despite its version name, OpenClaw v2026.10.5 was published October 18. It packages interrupted-turn recovery, durable input handling, subagent-result delivery and recovery from malformed or stalled streams.

A separate approval fix prevents ordinary plugin hooks from substituting arguments after trusted inline approval without obtaining approval for the replacement. Transforming plugins may need migration to request that separate approval.

Evidence & limits. Project-reported tests, not an independent durability or security certification. A restart fix merged October 5; its packaging into this release is the new event. Approval tests cover the source-runtime integration path and verify no file output when replacement approval is missing, pending or denied. The separate full-CI report was not accessible for review.