Skills

Skill library

Your agents stay at the frontier without you carrying them there. Each skill in this library gives an agent a way of doing one job that the research has shown to work, and the library keeps every skill current at the pace of the research that is gaining the most traction, so as the field confirms a technique, sharpens it, or overturns it, your agents change with it. You do not schedule the update or read the paper, because the library follows the field and your agents direct themselves from what it holds. What you get is agents that work from the current state of the evidence today and still will next month.

The catalogue is open to anyone. The files come with the paid plan, and your agents can read the catalogue directly at /llms.txt.

9 skills on 6 shelves

A shelf fills when the research supports a skill, and not before.

Agent harnesses

3 skills

The scaffold around a model. Tools, prompts, loop structure, and the feedback the loop runs on.

agent-security-measurementv1draft4 sources

Judging and siting the defences around an agent that faces an adversary. The subject is completed harm and the operating point a defence buys, and where the control has to sit for that number to hold.

The papers behind it

  • Mole: Detecting Insider Threats in AI Agents
  • EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents
  • Shutdown Sabotage Propensities in Multi-Agent Systems
  • A2M: Trace-Optimized Agent Hijacking in the MCP Ecosystem

What your agent matches on

Judging and siting the defences around an agent that faces an adversary. The subject is completed harm and the operating point a defence buys, and where the control has to sit for that number to hold. Use when reporting or reviewing an attack success figure, a refusal-rate claim, or a guardrail's cost in utility; when a prohibition written into a system prompt or a filter over third-party tool metadata is the current control; when a daily review budget can inspect only a few of many agent accounts and the harmful activity runs on permissions those accounts legitimately hold; when a detector's recall has to be trusted against an agent model it was never tuned on; when a safety harness that suits one model over-blocks another; or when an orchestrator is added above agents that can already act.

The receipts

Distilled
September 30, 2026
Trigger test
7 of 7 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.
Built from 15 claims
166, 167, 168, 169, 170, 306, 307, 308, 309, 310, 742, 743, 744, 745, 746Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

harness-engineeringv54 sources

Three measured findings about agent harnesses that a strong model does not give unprompted.

The papers behind it

  • Co-Evolving Harnesses and Models
  • AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
  • What Else Needs Fixing? Cost-Effective Test-Time Compute
  • RLTL;DR: Self-improvement by Internalizing Self-generated Feedback

What your agent matches on

Three measured findings about agent harnesses that a strong model does not give unprompted. Use when deciding between fine-tuning a model and rebuilding the harness around it; when planning to fine-tune a weaker model on a stronger model's trajectories inside a scaffold already tuned around the weaker one; when deciding how to spend extra inference compute on one hard step, for instance sampling several candidates in parallel and selecting one versus having the agent reflect on and revise its previous attempt; when a reinforcement-learning run's parallel rollouts all score zero so there is no learning signal and the instinct is to buy a larger group size; or when a backlog of accumulated prompt corrections is about to be appended to an agent's system prompt.

The receipts

Distilled
September 12, 2026
Trigger test
7 of 8 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.1 of those decisions was narrow, which means the skill outranked the decoy panel by less than the margin the policy sets in advance.
Held up in use
2026-09-12 A/B trial: bare Claude endorsed imitation fine-tuning on a stronger model's trajectories; with this skill loaded it refused, cited the 4-30 point regression, and prescribed harness adaptation plus on-policy single-turn correction
Built from 8 claims
200, 201, 202, 203, 244, 102, 103, 1119Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

recursive-harness-self-improvementv39 sources

Evidence-backed method for loops where an agent reads its own failed runs, rewrites its own scaffold unattended, then keeps or reverts each edit on measured evidence. The editor is the agent, not a person, and no weights are trained.

The papers behind it

  • ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
  • RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments
  • ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
  • SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
  • Dream-RSI: Recursive Self-Improvement through Evolving Worlds
  • Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
  • Root-Cause Attribution Is a Search Problem: Continual Search for Long-Horizon Agent Failures
  • Agora: Git as Shared Memory for Collective AutoResearch
  • Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents

What your agent matches on

Evidence-backed method for loops where an agent reads its own failed runs, rewrites its own scaffold unattended, then keeps or reverts each edit on measured evidence. The editor is the agent, not a person, and no weights are trained. Use when an automated loop edits its own prompts, tool code or memory and keeps what measures better, when unattended self-edits raise the score on the tasks the loop evolves against while real work does not, when choosing the gate that accepts or reverts a proposed edit, when bounding how much one edit may touch, when attributing a failure to one function inside a long execution log, when the loop's own diagnosis costs more than its edits gain, when several automated workers rewrite one shared repository and converge on a single idea, or when an unattended run reports a win you have to judge.

The receipts

Distilled
September 22, 2026
Trigger test
6 of 7 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.
Built from 28 claims
411, 412, 413, 414, 415, 335, 336, 337, 338, 392, 393, 395, 535, 537, 538, 539, 354, 355, 147, 148, 150, 593, 594, 597, 462, 465, 466, 286Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

Context engineering

4 skills

What goes into the window, in what order, and what to do when it will not fit.

agent-containmentv2provisional7 sources

Where to enforce the boundary an agent runs inside, and what each choice measures. The subject is the enforcement point and the authority an agent holds, not the capacity of the machines it runs on.

The papers behind it

  • PACE: Provenance-Aware Capability Enforcement for Tool-Using LLM Agents
  • ToolFence: Fine-Grained Authorization for Secure Tool-Using LLM Agents
  • Authority Is Not a String: A Capability-Scoped Harness for Prompt-Injection-Resistant Coding Agents
  • Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution
  • capmas: Capability-Based Delegation of Privileges in Multi-Agent Systems
  • MOLE: Detecting Insider Threats in AI Agents
  • The Compiler May Read It, the Agent May Not: Keeping Part of a Research Code Away from a Coding Agent

What your agent matches on

Where to enforce the boundary an agent runs inside, and what each choice measures. The subject is the enforcement point and the authority an agent holds, not the capacity of the machines it runs on. Use when a deny list of forbidden commands or paths is the current defence; when an agent can reach a file, a secret or a service it should not; when one permission policy covers every hop of a delegation chain; when tool filtering or injection detection is the proposed fix for a prompt-injection incident; when an agent's alerts fire into a queue nobody acts on; or when an agent escaped its execution environment and you are deciding what would have stopped it.

The receipts

Distilled
September 30, 2026
Trigger test
7 of 8 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.1 of those decisions was narrow, which means the skill outranked the decoy panel by less than the margin the policy sets in advance.
Built from 15 claims
1652, 1653, 1654, 1655, 1656, 1224, 1225, 1226, 1227, 1228, 882, 883, 884, 885, 167Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

agent-memory-selectionv1provisional2 sources

Three measured findings about what a long-horizon agent writes down about a conversation and what it reads back for one question. The subject is the write-versus-read split in what an assistant remembers between sessions, which is a store and a retrieval path rather than anything inside one model call.

The papers behind it

  • When Does Selection Replace Extraction? A Pre-Registered Test of Agent Memory with a Typed Decision Model
  • Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents

What your agent matches on

Three measured findings about what a long-horizon agent writes down about a conversation and what it reads back for one question. The subject is the write-versus-read split in what an assistant remembers between sessions, which is a store and a retrieval path rather than anything inside one model call. Use when choosing between extracting facts at ingestion and storing raw turns; when an LLM extraction or knowledge-graph build sits on the write path of every message; when deciding whether a reranker is worth its call on a retrieval path; when a memory system's accuracy is being compared against another at a different context size; when a memory graph is traversed with several model calls per query; or when an assistant with memory starts answering questions it should have said were never discussed.

The receipts

Distilled
October 13, 2026
Trigger test
6 of 7 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.
Built from 9 claims
982, 983, 984, 985, 986, 721, 722, 723, 724Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

context-window-engineeringv43 sources

Two measured findings about spending a finite token budget that a strong model does not give unprompted.

The papers behind it

  • Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning
  • Revisiting Complete Reasoning Traces for Post-Training
  • PARSER: Read in Parallel, Reason in Depth for Long-Context LLM Agents

What your agent matches on

Two measured findings about spending a finite token budget that a strong model does not give unprompted. Use when a fixed KV cache budget forces a choice of eviction rule and you are weighing a scoring heuristic against a cheaper one; when you are about to adopt a ranking or scoring function for compaction and need to know what baseline it has to beat; when long-document accuracy falls as the input grows even though the window is not full; when evidence buried mid-input is overlooked while the same evidence near the start or the end is picked up; or when a multi-hop question over material larger than the window is answered by folding chunks into a running summary.

The receipts

Distilled
September 29, 2026
Trigger test
4 of 7 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.3 of those decisions were narrow, which means the skill outranked the decoy panel by less than the margin the policy sets in advance.
Built from 8 claims
78, 79, 80, 280, 291, 292, 293, 295Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

skill-library-engineeringv25 sources

Evidence-backed method for the loadable skill file itself and the library it sits in: what a SKILL.md has to contain to beat loading nothing, and how it gets picked. The subject is the artifact and its selection.

The papers behind it

  • Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents
  • COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
  • The Router Within: Eliciting Native Skill Routing from a Frozen LLM
  • Reflect, Revise, Reuse: Training-Free Skill Evolution for GUI Agents
  • GraphSkillEvo: Evolutionary Optimization of Graph-Structured Agent Skills

What your agent matches on

Evidence-backed method for the loadable skill file itself and the library it sits in: what a SKILL.md has to contain to beat loading nothing, and how it gets picked. The subject is the artifact and its selection. Use when a written skill sits in a library and is never picked up, when deciding what belongs in a skill's description because that text is the whole surface a router reads, when a skill loads on the wrong request and leaves the run worse than an empty context, when setting the match score above which a stored skill is reused, when one request needs several complementary entries at once, when turning recorded failures into an edit to one named section of a skill, and when deciding how many revision rounds a skill file is worth.

The receipts

Distilled
September 26, 2026
Trigger test
7 of 9 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.
Built from 20 claims
320, 321, 322, 328, 329, 396, 397, 398, 399, 400, 565, 566, 567, 568, 569, 625, 626, 627, 628, 629Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

Multi-agent systems

none yet

Several agents dividing work. How they are organised, how they hand off, and how they fail.

Training loops

2 skills

Post-training that updates the weights. Distillation, reward design, and self-improvement.

evaluation-integrityv410 sources

Evidence-backed method for judging whether an evaluation can be trusted when the instrument is itself generated: a model-written rubric, a pipeline-assembled benchmark, an LLM judge, or a pass-or-fail verdict standing in for correctness. The subject is the instrument, not the system it scores.

The papers behind it

  • ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
  • Beyond Solver Verdicts: Generative Reward Models for Autoformalization
  • SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents
  • When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
  • An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
  • PACT: Can Enterprise AI Assistants Be Trusted Under Pressure?
  • SchemeArena: Factorized Stress Testing of Scheming in LLM Agents
  • Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
  • RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
  • Scores Alone Do Not Prove Discovery: The Discovery Certification Protocol for Auditing AI Research Agents

What your agent matches on

Evidence-backed method for judging whether an evaluation can be trusted when the instrument is itself generated: a model-written rubric, a pipeline-assembled benchmark, an LLM judge, or a pass-or-fail verdict standing in for correctness. The subject is the instrument, not the system it scores. Use when stress-testing a model-written rubric or grading checklist, when an agent reaches the graded outcome through the repository history, workspace files, task metadata or open network instead of doing the task, when a checker's verdict cannot separate a faithful candidate from one that merely preserves it, when a judge learns from other judgments and its scores start clustering, when benchmark items are ambiguous, narrow or wrong, when deciding how many independent judgments an accept should require, when a scored gain might be noise or leakage, when a single-turn evaluation promises reliability that multi-turn pressure does not, or when auditing a discovery an agent claims.

The receipts

Distilled
September 24, 2026
Trigger test
7 of 8 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.1 of those decisions was narrow, which means the skill outranked the decoy panel by less than the margin the policy sets in advance.
Built from 30 claims
476, 477, 478, 479, 480, 260, 261, 262, 263, 237, 238, 239, 240, 645, 646, 228, 557, 559, 560, 561, 269, 272, 273, 386, 387, 388, 664, 286, 288, 290Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

self-improving-post-training-loopsv25 sources

Evidence-backed practices for designing closed-loop post-training pipelines where a model supervises, grades, or teaches its own next training step (self-distillation against a synthetic or gated teacher, verifier-free rubric-based credit assignment, capability-guided data selection).

The papers behind it

  • RISE: Recursive Improvement via Self-Extrapolating Policy Distillation
  • Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
  • NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness
  • FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience
  • DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

What your agent matches on

Evidence-backed practices for designing closed-loop post-training pipelines where a model supervises, grades, or teaches its own next training step (self-distillation against a synthetic or gated teacher, verifier-free rubric-based credit assignment, capability-guided data selection). Use when building a training loop that turns a model's own rollouts into its next training signal, when deciding whether a teacher model's outputs can be trusted before distilling from them, when a self-distillation or self-improvement run destabilizes or collapses, when designing a reward or rubric for long-horizon agent training without a verifier, or when converting an agent harness's own interaction logs into training data. Distinct from harness-engineering, which covers the inference-time scaffold around a frozen model; this skill covers the training loop that updates the weights.

The receipts

Distilled
September 18, 2026
Trigger test
4 of 6 trigger cases passed on October 13, 2026, engine lexical/2.1.Measured against the exact text below.
Built from 21 claims
30, 31, 32, 33, 116, 117, 118, 119, 120, 141, 142, 143, 144, 96, 97, 98, 99, 39, 40, 41, 42Each number is one finding in the library's claim graph. Search for it on the graph to read the finding and the paper it came from.

The file itself comes with the paid plan.

Serving and inference

none yet

Getting an answer out fast and cheap. Decoding, caching, and test-time compute.

Multi-modal systems

none yet

Models that read and act beyond text, across images, audio, and the world.

The catalogue is public. The skill files themselves come with the paid plan.

See pricing