AI Agents

OpenComputer proposes a verifier-grounded framework for evaluating computer-use agents across 33 desktop applications and 1,000 tasks. Instead of relying on LLM-as-judge, it uses app-specific state verifiers exposing structured inspection endpoints. This approach aligns more closely with human adjudication, especially for fine-grained state changes where frontier agents often fail despite partial progress. The evaluation harness records full trajectories for auditable partial-credit rewards. This offers a concrete baseline for measuring robustness in GUI automation tasks.

PerfEvolve shifts database tuning from static documentation to dynamic agent actions. It translates expert tuning methodologies into executable skills for version-consistency verification and multi-parameter optimization. Evaluated on PostgreSQL under TPC-C and TPC-H, it outperforms documentation-driven baselines by 35.2%. The system handles inter-parameter dependencies that static guides miss. This demonstrates a clear path for devops agents: rather than retrieving docs, encode the tuning logic as executable skills that profile workloads directly. Open source integration is available.

TriMem addresses the limitations of atomic fact extraction in lifelong agent memory by maintaining three coexisting representation granularities. It stores raw dialogue segments for fidelity, extracted facts for retrieval, and synthesized profiles for deep reasoning. The system uses TextGrad-based prompt optimization to refine extraction without parameter updates. Experiments on LoCoMo and PerLTQA show consistent improvements over strong baselines. This architecture acknowledges that different reasoning tasks require different memory densities, avoiding the information loss inherent in compressing all history into static facts.

This paper defines the stochastic-deterministic boundary as the load-bearing primitive of production agent runtimes. It catalogs six runtime patterns, including hierarchical delegation and supervisor plus gate, tracing their lineage to distributed systems concepts. Crucially, it identifies replay divergence as a key failure mode where deterministic logs produce different outputs under model changes. A selection methodology covers coordination, state, and control concerns. This is essential reading for architects designing agent systems that must survive model variance and prompt drift.

SCARA targets vulnerability remediation in opaque industrial software where source code is unavailable. It connects binary vulnerability candidates to validated remedies through a four-stage pipeline, including operational-state-aware verification. On a 15-case benchmark spanning firmware and PLC artifacts, it achieved 100% precision with no false positives. Validation uses behavioral-coverage preservation and independent replay. This work proves agents can operate safely in high-stakes environments without source access, provided verification is grounded in operational state rather than static analysis alone.

Google released Gemini 3.5 Flash straight to GA, embedding it across Search, Android, and Enterprise platforms. The model supports 1M input tokens and 65K output, matching previous 3.x specs but skipping the preview phase. Notably, computer use capabilities are absent from this release. The new Interactions API is in beta, suggesting a push toward structured agent workflows rather than chat. Pricing details inform production cost decisions, signaling Google's intent to make this the default workhorse for high-volume agent deployments.

This alpha release of the llm-gemini plugin adds support for streaming reasoning tokens, compatible with llm >= 0.32a0. For engineers building CLI tools or local agents, this visibility into the reasoning process is critical for latency optimization and debugging. Streaming reasoning allows you to render intermediate thoughts before the final answer, improving perceived responsiveness. Necessary for anyone integrating Gemini reasoning capabilities into local development workflows where token cost and wait time matter.