AI Agents

The Kamera paper tackles a key inefficiency in multimodal agents: repeatedly re-encoding visual inputs like video frames or UI snapshots. It shows that naive, position-independent KV cache reuse breaks multi-hop reasoning by losing cross-chunk context. The proposed solution is a small, training-free, low-rank "conditioning patch" stored alongside each cache chunk. This allows for cheap reuse, reordering, and recall of cached context by simply re-rotating RoPE embeddings and applying the patch, restoring full accuracy without costly re-encoding.

This work identifies and names a crucial failure mode in agent learning: the "Self-Confirmation Trap," where an agent reinforces its own mistakes by misidentifying flawed but self-consistent trajectories as successes. To break this cycle, the authors propose EDV, an Execute-Distill-Verify framework. It uses heterogeneous agents for parallel exploration (Execute), a separate summarizer agent to reduce bias (Distill), and a consensus mechanism for validation (Verify) before committing any experience to memory, ensuring a more reliable learning process.

CodeTeam formalizes the "team of engineers" analogy into a multi-agent framework for repository-level code generation. It assigns specialized roles: multiple Architect agents propose software designs, a CTO agent evaluates and selects the best one to form a contract, Developer agents implement the code in a dependency-aware manner, and a QA agent runs tests. This separation of concerns improves planning and coordination for complex, multi-file software projects, outperforming monolithic approaches on benchmarks like SWE-Bench.

This paper moves beyond academic memory models to tackle the "fleet-memory problem" in production multi-agent systems. Based on learnings from a live service called MemClaw, it formalizes failure modes like unauthorized leakage, stale data propagation, and provenance collapse. To combat these, it defines necessary systems-level primitives: scoped retrieval, temporal supersession, and policy-governed propagation. The work emphasizes architectural insights from a real deployment, including discovered design flaws, making it a rare and valuable report from the trenches.

AutoSpec offers a way out of the trade-off between brittle hand-crafted safety rules and uninterpretable neural classifiers. The framework automatically evolves and repairs safety specifications for agents. Starting with expert-defined rules, it analyzes a stream of annotated execution traces, finds counterexamples (false positives/negatives), and then uses Inductive Logic Programming (ILP) to generate candidate rule edits. This produces interpretable, refined rules that better match observed safe and unsafe behavior in the wild, balancing precision and recall.

Holmes is a multi-agent system for root cause analysis of mobile app crashes, demonstrated at industrial scale on the 70-million-line WeChat codebase. Because crashes are often non-reproducible, it doesn't try to rerun them. Instead, it synthesizes multimodal signals from the post-mortem diagnostics—stack traces, logs, thread states—to reconstruct the failure context. The system uses a Retrieve-Explore-Reason architecture to navigate the massive codebase, achieving 87.6% accuracy and reducing human investigation time by over 98% to about 77 seconds.

Coding agents waste huge budgets just finding where to apply a fix. SHERLOC is a training-free framework designed specifically for structured fault localization. It pairs a reasoning LLM with compact repository tools to generate not just a location, but also diagnostic context for a separate repair agent. On SWE-Bench, it sets a new state-of-the-art for localization, and when its output is piped into repair agents, it boosts their success rate by nearly 6 percentage points while reducing token consumption.

This paper tackles the thorny problem of evaluating agentic data analysis systems, asking "how do you reliably grade the grader?" The authors propose and evaluate a three-layer cascade for scoring complex outputs: strict regex matching, a more forgiving LLM-based grader, and final human inspection. They found the automated graders could achieve 100% precision, with the LLM reaching 97% recall. The study also shows how an "iterative nudge" mechanism can coach the agent into producing more easily parsable output.

This paper presents a well-designed head-to-head comparison of GUI-based and CLI-based agents. On a benchmark of 440 desktop tasks with identical goals and verifiers, a screen-only GUI agent achieved a 59.1% pass rate, beating a CLI agent with its initial skill set (48.2%). However, once the CLI agent's skills were augmented, its success rate jumped to 69.3%. The study concludes that the two modalities have different bottlenecks: GUIs are limited by interaction reliability, while CLIs are limited by skill coverage.

Instead of using fixed rules for tool orchestration, this paper proposes a more principled approach using Bayesian control. It frames the coding agent's workflow as a cost-sensitive sequential decision problem. A Bayesian controller maintains a probability distribution over a code candidate's correctness, and at each step, it dynamically decides whether the most cost-effective action is to gather more evidence (e.g., run a cheap linter), refine the code, run an expensive verifier, or stop. This optimizes the trade-off between verification and cost.

Simon Willison explores a new frontier for browser-based agents: persistent storage. He's investigating if the Origin Private File System (OPFS) can allow Python apps running in the browser via Pyodide, like Datasette Lite, to edit local SQLite files. This would enable stateful, local-first applications without a server backend. To that end, he built a simple test harness with assistance from a code agent to probe browser support and feasibility. The implications for local agent execution are significant.

The latest Datasette alpha (1.0a35) ships major features for programmatic data interaction, highly relevant for agent tool use. The release introduces new "Create table" and "Alter table" actions, both backed by corresponding JSON APIs. This means an agent can now not only query data but also define and modify database schemas, including columns, primary keys, and foreign keys. This turns Datasette into a more powerful and stateful tool for data-centric agents to wield.