Shows what happened
Useful traces do not decide what should guide the next run.
Agent Experience Intelligence · 5-minute brief
Verified experience for the next agent run.
AEG turns real execution into reusable guidance—and tests whether that guidance actually improves the next task.
01 / Repeated-learning problem
Every new run starts too close to zero.
Teams keep code, traces, and documentation. They rarely keep a trustworthy account of which recovery worked, under what conditions, and whether reusing it helped.
Says what should happen
General guidance rarely carries run-level outcome evidence.
Tests what transfers
Unlike generic memory, AEG keeps provenance and a verified outcome attached, then tests whether the experience should guide the next run.
02 / Operating loop
Turn execution into evidence, then test the reuse.
The cross-platform layer sits beside agents, models, tools, and observability. It captures a run, sanitizes the reusable signal, verifies the outcome, retrieves relevant experience or abstains, and measures what happens next.
Run receipt
Attempts, failures, recovery, and outcome.
Safe record
Portable fields and sanitized content.
Evidence
Oracle, provenance, applicability, limits.
Guidance
Explain relevance—or correctly abstain.
Comparison
Success, commands, tests, tokens, latency.
03 / Unit of learning
The experience record preserves the path, not just the answer.
A verified record is situated evidence. It is useful only when its provenance, applicability, and limitations remain attached.
04 / Initial wedge
Start where transfer can be tested cleanly.
The first tasks are reproducible failures with an objective success check, comparable baseline and assisted runs, and low privacy or IP risk.
Engineers using coding agents
Real repositories, bounded failures, objective verification.
AI dev-tool and OSS maintainers
Repeatable task families where recovery knowledge gets lost.
Enterprise AI-platform teams
Governed reuse across internal agent workflows—if transfer holds.
Long-term thesis
From Superintelligence to Distributed Intelligence
The future may not be one agent knowing everything. One possible topology is billions of human-agent systems learning locally and sharing selectively—when evidence, permission, and context support the transfer.
General capability
What can an intelligent system broadly do?
Current task knowledge
What does this task need to know now?
Verified situated experience
What worked before under comparable conditions?
- attempts, failures, and recovery
- verified outcome and cost/latency
- provenance and applicability
05 / Shipped and measured
Capture, validation, and retrieval are shipped. Benefit remains bounded.
Repository evidence keeps shipped facts, positive results, negative results, and unanswered questions separate.
5 pairsone repair family
10 / 10arms passed
1 fewermedian completed command
+18.2 smedian assisted latency
Median assisted runs used one fewer completed command. This bounded five-pair result did not improve success, and latency regressed.
| Claim | Status | Repository evidence | Boundary |
|---|---|---|---|
| AEG can capture, validate, and retrieve experience. | Shipped | VS Code v0.1.5, schemas, validator, recommender | Developer preview; local-first |
| Verified records can preserve auditable outcomes. | Verified |
Two records in experiences/registry.json
|
Outcome evidence ≠ causal AEG benefit |
| A bounded repair run used fewer median commands. | Bounded | Five pairs; all ten arms passed objective verification | No success gain; median latency regressed 18,235 ms |
| Related transfer pairs produced positive retrieval evidence. | No | One neutral pair; one preregistered result not positive | Negative and neutral evidence stays visible |
| The autonomous loop follows bounded state and approval rules. | Validated | Hash-chained transitions and approval-gate stop | Workflow validation only; not retrieval benefit |
06 / Evidence boundary
What the evidence does not say.
- No demonstrated success-rate improvement
- No reliable cross-project learning claim
- No proven reduction in retries, tokens, cost, or latency
- No customer adoption or product-market-fit evidence
- No global reputation, popularity, or safety scoring
- Not a generic marketplace or agent orchestrator
- No LangSmith replacement claim
07 / Current market status
Stage A outreach is authorized. The public budget remains at zero.
Under the merged Stage A approval record, individually reviewed outreach is authorized for up to three voluntary seed participants. The public recruitment budget currently records 0 invitations and 0 enrolled participants. Task execution and AEG-assisted testing remain unauthorized.
08 / Next falsifiable milestone
Earn one controlled transfer result in a real workflow.
Recruit a small seed cohort, accept only reproducible low-risk tasks, then compare baseline and AEG-assisted runs with retrieval as the intentional difference.
Decision rule
Relevant prior experience must improve at least one preregistered efficiency measure—retries, commands, tests, tokens, or time—without lowering objective task success. Correct abstention, neutral results, and regressions remain evidence.
Design-partner experiment
Bring one reproducible task.
Bring a reproducible task that is public and license-compatible, explicitly authorized, or synthetic, with an objective success check. Any retained evidence will be sanitized. AEG will run a bounded baseline-versus-experience test and return an auditable result.
AEG contributes experiment design and implementation, the comparison, a sanitized record, and a clear result with limits.
09 / Technical architecture
A cross-platform execution-evidence layer.
AEG complements agent runtimes and observability. It connects execution inputs to governed experience records, retrieval with explanations or abstention, and measured outcomes.
Coding agents & IDE workflows
Tasks, traces & test artifacts
Developer feedback
AEGverified experience layer
Evidence schema & provenance
Retrieval, explanation & abstention
Outcome comparison & recapture