Better agents need
explicit state.

A memory system connects personal facts, reusable procedures and the links between them. When a fact changes, the procedures depending on it can change too.

The authors report 40.78% session success versus 20.60% for a Claude Code native-memory baseline, with the same Qwen3.7-Plus agent across methods. This is a memory-method comparison, not a comparison against the stock Claude product.

Evidence & limits. 20 simulated users and 500 sessions; one full inference run per method. The results are promising, but simulation and single-run evaluation limit generalization.

HPD turns reference trajectories into runtime-control documentation, then verified action-and-control training examples. At deployment, the agent uses the native interface without the privileged documentation.

On 583 held-out tasks, Qwen3-8B task success rises from 3.09% to 22.98%. The strongest OPD comparison reaches 15.61%.

Evidence & limits. HPD is not best on every metric: compliance is 49.23%, versus 66.21% for OPD. It uses a harness-specific adapter and does not train on error-recovery states.

Claude adds OpenTelemetry support for the Messages API and Managed Agents across seven SDK languages. Metadata-only export is the default.

The release provides a standard path for instrumentation. Full-content export can include prompts, responses and tool content, making it a separate privacy decision.

Evidence & limits. Restricted organizations cannot export content. This is a public beta: schemas may change, so pin SDK versions and check the telemetry contract.

OpenCompute connects self-hosted agent sandboxes through an outbound connection. Drivers manage lifecycle operations, with optional suspend, snapshot and fork capabilities and conformance checks.

The architecture separates the sandbox’s execution environment from the agent service, which can make infrastructure policy more explicit.

Evidence & limits. Access is limited. Open source is planned, not already available. Local execution does not establish local model inference or local session storage.

The release addresses six security advisories affecting scenarios including Docker deployments handling untrusted requests: SSRF, script injection and denial of service are among the issues.

Maintainers recommend upgrading. Exposure depends on deployment configuration and dependencies, rather than applying identically to every installation.

Evidence & limits. The remote-code-execution issue tied to PyTorch versions below 2.6 does not affect the default image without PyTorch. Availability of an updated Docker image was not verified.

LangChain describes tool-output offloading at 20K tokens, human-message offloading at 50K, compaction at 85% of the context window, and retrieval from an archive. It also covers skills, subagent isolation and plain-file memory.

Useful implementation guidance for managing context as a resource. These are the guide’s thresholds and mechanisms, not universal optimal settings.

Evidence & limits. The publication is a new guide; it does not mean every described feature is new. No comparative benchmark is supplied.

An outer harness inspects unfinished Codex trajectories and chooses retry, termination or human escalation. The case also describes reducing a backtest from three hours to three–five minutes.

Faster feedback reduces repeated polling and context pollution. The interesting pattern is the separation of execution from supervisory control.

Evidence & limits. This is an OpenAI vendor/customer report about historical work, not an independent benchmark or an event necessarily occurring on publication day.