AI Agents

EVE-Agent: Evidence-Verifiable Self-Evolving Agents

Current reflection methods only self-correct within single tasks. This paper introduces In-context Training, where a reflector model observes actor trajectories and generates system prompts to improve future unseen tasks. An RL pipeline learns reflections without human examples. Trained reflectors outperform baselines on ALFWorld and MiniHack, with some generalization to different environments. MetaGym library enables future research.

Enterprise skills need more than task guidance—they require inspectable goals, input boundaries, permissions, evidence requirements, output contracts, and human approval points. This GovernSpec-inspired framework organizes SKILL.md files as readable task contracts. Evaluation shows contractual skills outperform no-skill baselines across models, with gains mainly in checkability and maintainability rather than raw generation quality.

Long-horizon agents accumulate histories exceeding context windows. LLM summarization is lossy and blocks inference for seconds, with unpredictable output volume. Parallel compaction gives operators fine-grained control over summary volume and enables targeted prompt engineering per block. Evaluated across four backbones from 8B to 120B parameters, it reduces wall time and improves throughput at matched decode volume versus sequential baselines.

Agent benchmarks fragment across success rates, tool validity, trajectory safety, and attack robustness. AgentAtlas proposes a six-state control taxonomy, nine-category trajectory failure taxonomy, taxonomy-aware versus blind methodology, and benchmark coverage audit. Testing eight models shows removing explicit label menus drops trajectory accuracy 14-40 percentage points to a 0.54-0.62 floor regardless of model family.

This HN-discussed paper analyzes failure modes when LLM agents generate backend code. The title suggests agents lose constraint adherence over longer generation tasks, creating reliability risks for production deployment. With 246 points and 143 comments, the community recognizes this as a critical safety concern for code agents operating in real systems.

A community-built DeepSeek wrapper optimizes for production costs through aggressive caching. The thread notes DeepSeek made V4 Pro pricing discounts permanent, signaling competitive pressure on inference costs. 585 HN points suggest strong developer interest in cost-effective coding agents for production workloads.

An opinion piece argues against delegating architectural decisions to Claude. The HN discussion reflects ongoing tension between agent capability claims and actual system design responsibility. Lower engagement suggests the community wants concrete engineering data over positional arguments about agent limitations.

Datasette-agent 0.1a4 integrates the agent chat interface directly into Datasette's Jump menu, triggered by hitting /. Users initiate conversations that execute data operations like counting entries, with results returned inline. The demo runs on agent.datasette.io with GitHub authentication. This represents a practical tool-use pattern where the agent operates within an existing data exploration workflow rather than as a standalone chatbot.

David Oks explains how AI datacenter demand is reshaping memory economics. HBM now claims 20% of wafer capacity versus 2% previously, and each HBM gigabyte consumes 3x the wafer space of DDR or LPDDR. With only three major memory manufacturers preferring under-provisioning, consumer device RAM faces multi-year constraints. Budget smartphones in emerging markets feel this first.

Armin Ronacher documents a production failure mode: AI-generated issue reports that lack human voice and contain confident but inaccurate root cause analysis. These slop issues include fake minimal repros and wrong code analogies. He requests issue reports condense to: what command ran, what was expected, what happened, and the exact error log.