Ready to run agent loop
Demo ready
New agent task
Build a documentation ingestion pipeline for an agent coding assistant.
Click Run 45-second loop demo
Watch a new task become an experience-guided execution loop.
BrowserBC packages reusable skills. SkillOpt improves skills from feedback. Loop engineering turns agents into continuous plan-act-verify systems. AEG supplies the execution intelligence layer: it recommends the right experience package and updates the loop after every run.
Watch a new task become an experience-guided execution loop.
Loop engineering is the operating pattern. AEG is the experience layer that improves each iteration: what to try, what to avoid, when to stop, and how to validate.
Prompt libraries and RAG retrieve content. BrowserBC-like systems package domain skills. SkillOpt-like systems improve skills. AEG retrieves and ranks the full execution experience needed for the current loop.
Store workflows, skill cards, optimized skill versions, loop policies, failure modes, recovery plans, validation signals, and outcomes as one reusable object.
Recommend experiences that materially improve success, reduce retries, lower token cost, or prevent runaway loops.
Every loop updates skill confidence, failure patterns, benchmark evidence, and future recommendation scores.
AEG does not compete with skill libraries, optimizers, or agent runtimes. It makes them more useful by retrieving the right experience at the right time.
Human browser paths become structured reusable skills with intent, steps, recovery, and constraints.
Execution feedback improves skill instructions and selects stronger skill versions over time.
Agents plan, act, observe, verify, refine, and repeat until the task is complete.
Given a new task, retrieve traces, workflows, skills, failures, guardrails, validation gates, and loop policies.
AEG is not another agent framework. It is an execution intelligence layer that can sit across agent loops, tools, skill libraries, optimizers, and evaluation systems.
Define sanitized trace + workflow + skill + failure + validation + loop policy objects.
Recommend experience packages by similarity, success evidence, risk, and expected cost savings.
Suggest retry budgets, stop conditions, validation gates, and human-in-the-loop checkpoints.
Import BrowserBC-style skill cards, SkillOpt-style versions, MCP tools, and agent runtime traces.
Share reusable loops, benchmarks, failure cases, and validated skill packs across the community.
Agent Experience Graph is an early open-source exploration of reusable execution knowledge for AI agent loops.