AI Agents

Instead of building environments for specific tasks, this framework instantiates persistent worlds — entities, services, tools, state, and executable cross-service invariants — from high-level business scenarios, then lets diverse tasks emerge from the world. They generate 4,783 executable environments across 14 industries and use them as RL training substrate. Qwen3.5-4B improves 12.3→15.7 on EnterpriseOps-GYM and 45.9→56.0 on AIME26 without targeting those benchmarks, and the construction process itself is learnable. The out-of-domain generalization to reasoning and coding is the striking result.

A hierarchical memory system that organizes execution into tasks, subtasks, and actions and assigns each memory a dynamic retention score, with event-based updates and selection-based decay deciding what stays active versus gets folded away. On GAIA-Text across three small models it beats linear memory by 9.97pp accuracy while cutting prompt tokens 32.8%, and memory-poisoning experiments show it limits propagation of unreliable content. The thesis — long-horizon memory is about what to keep active, not how much to store — is well supported here.

An interventional benchmark for causal failure attribution in multi-hop agentic RAG: inject a certified fault at a chosen hop, re-execute downstream, and test whether diagnosers can still identify the injected hop after the suffix changes. The headline finding is that coverage-based diagnosis scores 0.91 at hop 1 but 0.00 at hops 2 and 3 — post-hoc traces lose the signal once failures propagate. A counterfactual probe recovers some signal (0.67) but on tiny samples (18, then 3 cases). Propagation depth as an explicit eval axis is the real contribution.

A repository-native framework for evaluating skills and capability packages as executable artifacts: run paired live trials with and without a target skill under the same model, sandbox, and grader, then report Skill Lift — the added value for a fixed task. On 145 real skills, structure-scanning gates correlate weakly with LLM-judge outcomes (Spearman ρ=0.14), confirming they measure different things. Across 947 paired cases, mean composite lift is 0.213 with a tight CI, positive in 72.8% of pairs. Concrete production data, not a toy study.

Runtime intervention usually needs an expert solver or a task-capable critic to generate corrections — both expensive. This work instead trains a tiny comparator that only judges whether sampled alternatives beat the actor's proposal, using pairwise supervision from same-prefix counterfactual branches; preferred alternatives are returned as non-binding advice. Across WebShop, ALFWorld, and tau^3-Retail with three actors, it improves all nine settings and beats baselines even though the advisor is substantially weaker at the task than the actor it advises. Comparison, not solving, is the trick.

A microharness for persistent agents, posted to HN with modest traction (68 points). The pitch is a minimal scaffold for keeping agents running across sessions rather than a heavyweight framework, but the writeup is thin on the actual persistence mechanism — how state is checkpointed, restored, and bounded over time. Worth a look if you're building long-lived agents and want something smaller than the incumbent frameworks, but read the code before adopting.

A practitioner essay drawing a line between the agent harness — the loop, tools, memory, retries — and the underlying model, arguing that most of what people attribute to model quality is actually harness engineering. It's essay-level depth, not measured, but the framing is a useful corrective for teams that keep swapping models to fix behavior that lives in their orchestration layer. The comment thread (34 replies) suggests the distinction resonates with people shipping agents.

Microsoft's agent-lightning hits v1.0 GA. The project targets training and optimizing agents — including RL-style loops — decoupled from the agent's application code, so you can improve behavior without rewriting the harness. This is the milestone release with a version number and code, but the HN post is terse (51 points, 6 comments) and the release tag alone doesn't convey what changed since the pre-1.0 line. Check the changelog for API stability guarantees before you build on it.

A widely-shared writeup (404 points) of one engineer's agent.md config for steering LLM coding assistants toward higher-quality output — conventions, constraints, and repository context the agent reads before generating. It's fundamentally a tips-and-config post, not a technique with measured lift, but the popularity signals that curated per-repo instructions remain the cheapest reliability lever available. Treat it as a starting template to fork, not a benchmarked method.

This paper characterizes a real abuse mode: agents flooding government services — form submissions, requests, queues — at machine scale, overwhelming systems built for human throughput. It names and measures the phenomenon rather than proposing a fix, which is the right first step for a threat that most operators aren't yet modeling. The lively comment thread (79 replies) reflects genuine disagreement about attribution and defense. Useful if you run public-facing services that assume human-rate traffic.