Start from zero
Similar tasks are decomposed, explored, and debugged again—without access to prior execution lessons.
Agent Experience Intelligence
Turn real agent runs into verified, reusable experience.
AEG captures what agents tried, what failed, what recovered, and what worked—then turns those runs into evidence-backed guidance for future tasks.
The problem
Agents keep paying the same learning cost.
Execution experience disappears across sessions: useful failures, recovery steps, cost, and validation are rarely captured in a form the next agent can trust.
Keep the answer, lose the path
Attempts, dead ends, recovery steps, latency, and token cost vanish even when the final patch survives.
Guidance lacks evidence
A plausible tip is not a verified experience. Agents need provenance, constraints, outcomes, and reproducibility.
The operating loop
Make execution evidence useful at the next decision.
AEG focuses first on coding agents. It is an evidence layer—not a generic agent marketplace, an orchestrator, or a replacement for observability platforms.
Record the run
Intent, context, steps, skills, artifacts, failures, recovery, outcome, and cost.
Shape the signal
Sanitize and structure the reusable lesson without raw private work.
Test the outcome
Attach objective checks, limitations, and reproducible provenance.
Explain the match
Retrieve relevant lessons and show why they apply—or abstain.
Guide the next run
Offer a guarded capsule, then capture the new outcome and feedback.
Evidence, not theater
What exists. What is verified. What is still a hypothesis.
AEG’s public artifacts are useful precisely because their boundaries are visible.
Working developer surface
- Local-first VS Code extension v0.1.5
- Open trace schema and validator
- Explainable recommender and demo
- Two records in the verified library
Auditable, narrow results
- Five paired trials on one FastAPI repair family
- All ten arms passed with the same fix
- Median assisted runs used one fewer completed command.
- Median assisted wall time regressed by 18,235 ms
Does retrieval improve the next run?
Test whether relevant prior experience reduces retries, token usage, and execution time without lowering task success.
5 pairssingle repair family
10 / 10arms passed verification
1 fewermedian completed command
+18.2smedian wall-time regression
Source: public repair lab results. This does not establish success-rate improvement, generalized transfer, or product-market fit.
What AEG preserves
A compact receipt for agent work.
The reusable unit contains enough evidence to judge a recommendation without retaining raw conversations or proprietary code.
↳
Execution
Intent, context, decomposition, attempted skills and tools.
!
Recovery
Failures, rejected paths, guardrails, and the steps that recovered.
✓
Verification
Outcome, objective oracle, provenance, reproducibility, and limitations.
∑
Cost
Attempts, commands, tests, latency, and tokens when reliably observable.
Next validation milestone