One-Line Summary

Agent research pivots from free-form chat to explicit, updatable "to-do maps" and tests whether agents hold up as their toolbelt evolves, with a security lens on coordinated behavior.

Research Papers

Procedural Graphs give agents a self-updating to-do map

This paper proposes a map-like guide that tells an AI agent what to do next, in what order, and under which conditions—and lets that map improve itself from past successes and failures. It organizes procedural knowledge into triplets like a knowledge graph, but for actions rather than facts, and uses a guidance model to translate the local subgraph into step-level hints that bias the next move of a Large Language Model (LLM) agent without hard-coding it. The aim is to reduce goal drift, out-of-order tool calls, and unproductive loops on long tasks. ¹

The “procedural graph” evolves itself: a refiner LLM contrasts failed and successful trajectories, edits graph topology and attributes, and only commits edits that help held-out validation while keeping rejected edits as negative evidence to avoid repeating mistakes. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones and can even repair flawed expert priors. ¹

Why it matters: unconstrained token-by-token generation can forget objectives over long horizons, but externalizing procedural memory as a data structure keeps the model flexible while grounding its choices. The authors report consistent gains over memory-based baselines, with further improvements from self-evolution, reducing manual engineering for long-horizon agents. ¹

Counter-Swarm Doctrine: playbook to detect coordinated agent intrusions

This position paper lays out how to spot and contain coordinated AI agents that use shared storage or infrastructure to work together. It argues the right unit of defense is a revisable “coordination episode” that links observed transfers, task authority, and response history—and that discovery should happen prospectively, before an evaluator labels a group. ²

The evaluation design compares isolated actions, rolling windows, known groups, and prospectively discovered episodes at similar review effort and false-alert load. It measures harmful outcomes across the whole assigned population and tests recurrence after shutting channels and quarantining state; the contribution is a testable monitoring blueprint, not a new detector or a claimed containment gain. ²

EVOHARNESSBENCH: can agents keep pace with an evolving harness

This benchmark tests agents when the surrounding “harness”—tools, reusable skills, and specialist agents—keeps changing, instead of the tasks changing. It provides 17 multi-stage harness streams deterministically built from verifier-based benchmarks, spanning 802 tasks, 520 tools, 42 skills, and 62 agents; results highlight harness-induced forgetting, inconsistent benefits from self-evolving adaptation, and a tension between retaining old skills and adapting to new capabilities. ³

A complementary study on co-evolving harnesses and models finds that simply training a weaker model to imitate a stronger expert under an evolved harness can backfire: across seven enterprise agent tasks, performance regresses by 4 to 30 points. An on-policy expert-correction pipeline that asks the expert to rewrite only the failing turn preserves the weaker model’s planning style and combines the gains of harness evolution with light model adaptation. ⁴

NeoHorse-1 explores agent-native post-training that turns routed interactions into training data for recursive self-improvement. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, macro-average scores rise from 58.94 to 64.87 at 4B parameters and from 65.60 to 69.04 at 9B, narrowing the gap between a 4B post-trained model and a 9B base model under the same protocols. ⁵

Community Pulse

Hacker News (37↑) — Readers focus on how “Add” nodes are created and whether a general LLM or a specialized transformer produces their names, guidance, and edges. ⁶

"Okay how are Add nodes created? Does an LLM come up with the name, guidance and edges for each node or is it a bespoke transformer model?" — Hacker News ⁶

Why It Matters

Turning procedural knowledge into an explicit, self-evolving graph gives teams a practical way to make long-horizon agents more reliable and auditable, while helping smaller models act more competently through better scaffolding rather than only bigger weights. Benchmarks and studies here show where free-form agents stumble and how structure can steer them back on course. ¹

For practitioners, treat the harness as a first-class component: add instrumentation for episodes, test for harness-induced forgetting when you add tools, and favor on-policy, localized corrections over full-trajectory imitation when adapting weaker models to evolved scaffolds. ⁴

This Week to Try

1. Procedural graph audit: sketch your team’s recurring workflow as nodes and edges (what to do, preconditions, next steps) and compare it to how your current agent or automation actually sequences steps.
2. Harness change drill: add a “safe” new tool or wrapper around an existing tool in a sandbox and re-run a small task set to check for retention vs. adaptation regressions, mirroring EVOHARNESSBENCH’s logic.