Sentient's EvoSkill Auto-Discovers Agent Skills Through Evolutionary Loops

Tom Dörr

Mar 13, 2026 · Updated Jun 6, 2026

Sentient AGI released EvoSkill, a self-improving agent framework that automatically discovers high-performance skills for AI agents. It runs an evolutionary loop that tests agents on benchmarks, identifies failure patterns, and proposes improvements without manual prompt engineering.

EvoSkill, a self-improving agent framework by Sentient AGI, automatically discovers high-performance agent skills through an evolutionary loop. The framework treats agent configurations as versioned programs and iterates through five stages: a base agent runs benchmark questions, a proposer analyzes failures and suggests changes, a generator writes new skill files or prompt rewrites, an evaluator scores the variant on a held-out set, and a frontier tracker preserves the top performers as git branches.

Manually crafting agent skills is a major bottleneck in building reliable multi-agent systems — it requires deep expertise and constant iteration. EvoSkill removes that constraint by automating the optimization cycle entirely. The framework has been validated on DABStep, SEAL-QA, and OfficeQA, showing automated discovery can match or exceed hand-tuned configurations.

Point it at your own benchmark tasks and let the loop replace manual skill iteration — no prompting expertise required. Both Claude and opencode SDKs are supported.

View the full update on github.com

Tom Dörr

@tom_doerrMar 10

Synthesize agent skills via evolutionary loops https://t.co/vRq9PSmJm3 https://t.co/KnoBlwUHTm

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Keep reading

Anyscale Agent Skills Gives Coding Agents Native Expertise for Ray Workloads

Anyscale launched Agent Skills in general availability to equip AI coding agents with specialized knowledge for building and debugging distributed Ray workloads. By encoding infrastructure constraints and version-aware APIs into the agent loop, the system prevents common failures like out-of-memory errors.

Minko GechevMar 26

Skillgrade Brings Regression Testing to Agent Skills via Automated Evals

Skillgrade released an open-source CLI that runs automated evals against agent skills, catching regressions when a skill, model, or agent changes. Until now, there was no standard way to verify agent skills hold up across model updates.

AWS Releases Strands Evals to Systematically Test Non-Deterministic AI Agents

Amazon Web ServicesApr 2

AWS Releases Strands Evals to Systematically Test Non-Deterministic AI Agents

AWS introduced Strands Evals, a framework that uses LLM-based judges and multi-turn simulations to evaluate AI agents. Unlike traditional software testing, this system measures non-deterministic behaviors like helpfulness, tool accuracy, and goal success. It provides a structured path for moving agents from experimental prototypes to reliable production deployments.

Elastic DevMar 18

Elastic Open-Sources Agent Skills to Give Coding Agents Platform Expertise

Elastic released Agent Skills, open-source instruction packages that give AI coding agents curated expertise for Elasticsearch, Kibana, Observability, and Security. Skills load into any runtime — Cursor, Claude Code, Copilot — replacing guessed syntax with correct, version-aware operations.