AI Agents

Dynamo is a new training-free framework that lets a frozen vision-language model evolve its own skills and tools. By inspecting its own correct and incorrect attempts on a small set of labeled examples, the agent builds a persistent library of reusable reasoning patterns and executable visual tools. This process of self-correction and accumulation improves accuracy on five different VLM backbones across four visual reasoning benchmarks by an average of 5.6 points, closing 65–99% of the gap with expensive task-specific RL tuning.

AgentFlow is a static analysis framework designed to understand agent programs that mix conventional code with framework-specific semantics. It builds an Agent Dependency Graph (ADG) to represent relationships between agents, prompts, models, and tools that are often implicit in framework APIs. By making these dependencies explicit, AgentFlow can generate an "Agent Bill of Materials" for governance and perform security analyses like tracing data flows from prompts to risky tool calls. The tool supports five major agent frameworks and was evaluated on 5,399 real-world programs.

This work introduces Janus, an open-source playground and evaluation harness for designing and testing user-involved permission systems for agents. Grounded in a conceptual model of user involvement, the authors implement and test six different "permission assistants" across three scenarios. The results confirm that user input is critical for security, AI-augmented decisions can reduce user cognitive load, and "permission fatigue" is a real failure mode that must be designed for. No single permission design was optimal, arguing for context-sensitive approaches.

This paper challenges a common assumption in code generation: that ambiguous prompts produce diverse, incoherent code samples. The authors identify "detrimental semantic collapse," a failure mode where models consistently generate the same incorrect program from an underspecified prompt. This behavior was found in over 10% of MBPP tasks and 32% of LiveCodeBench. This finding undermines evaluation and disambiguation techniques that rely on sampling multiple outputs and using their variance as a proxy for prompt ambiguity or model uncertainty.

ContextNest proposes a formal governance layer to sit beneath RAG systems, focusing on the provenance and integrity of retrieved knowledge. It's an open specification for vaults of AI-consumable knowledge, using contextnest:// URIs, hash-chained version histories, and cryptographic checkpoints. This allows an agent's context to be verifiably reconstructed for audit. In a stale-version attack experiment, this governance-first approach had a higher pass rate (97% vs 90-93%) at one-third the token cost compared to standard BM25 retrieval, which failed to identify the correct document version.

Armin Ronacher reports a counterintuitive failure mode: newer, more powerful Claude models are worse at using a third-party edit tool than their predecessors. While Opus 4.8 and Sonnet 5 produce the correct edit logic, they frequently add extra, invalid fields to the tool call arguments, causing the call to fail. The hypothesis is that RL training on Claude's own internal tools has caused negative transfer, degrading performance on tools with slightly different schemas. This is a sharp reminder that model upgrades can introduce subtle regressions.

Simon Willison provides a concrete case study on using Claude Fable for a pre-release code review of his sqlite-utils library. With a single prompt asking for a final review to spot breaking changes, the agent identified five "release blockers" that Willison had missed. The most critical was a data loss bug where Table.delete_where() failed to wrap its DELETE in a transaction, leaving the connection in a poisoned state. The entire review process cost about $150, demonstrating a practical application of agents for code quality assurance.