“We already fixed this.”
Repeat the correction.
YOUR CODING AGENT'S LEARNING LAYER
You fixed it.
Why teach it
again?
SenseLab turns verified fixes into learning your agents can reuse. Across sessions, tools, and teammates.
Your IDE stays. Your models stay. The experience carries forward.
YESTERDAY / YOU + CLAUDE CODE“Use event.id. We fixed this.”
verifiedHandle the same event, retried 3 times.
Yesterday's fix didn't make the handoff.New session. Same wrong assumption.
12
const key = delivery.requestId;LOCAL RETRY TESTsame event × 3
evt_42+ jobevt_42+ jobevt_42+ job
3
jobs. One event.The same bug got another shift.
FAILReal local test. Agent behavior illustrated.
Sequential browser fixture, not production deduplication. Production needs persistent, atomic handling. No model or SenseLab API calls.
ALREADY USING A CODING AGENT?Keep it. Add SenseLab.
Codex + SenseLab 01 / THE PART YOU DIDN'T AUTOMATE
AI writes the code.
You still carry the context.
“Let me paste the context.”
Rebuild the backstory.
“My agent didn't know that.”
Redo the investigation.
02 / MAKE THE EXPERIENCE USEFUL
Less re-explaining.
A better starting point.
Scroll to follow one fix. Try each step.
The fix shipped.
The lesson didn't.
It's in a chat. Or a growing .md file.
The next agent still needs the right context.
Keep the fix.
And why it worked.
Verify it. Record the finding, its scope, and the outcome in SenseLab.
FindingEvidenceOutcome
New task.
Not a blank slate.
Retrieve the relevant lesson before coding. Even in a different tool.
Now give the team a head startwebhooks / retry handlingsession 01
NO LESSON RETRIEVEDClaude Code
Same event. Three jobs.
01
for (const delivery of retries) {02
const key = delivery.requestId;03
if (!seen.has(key)) createJob();04
seen.add(key);05
}
THE MISSING HANDOFF“Use event.id, not requestId.”A previous correction. Not consulted here.
×
Expected 1 job. Got 3.Same event. Different request IDs.
Retrieve→Work→Report outcome
Executable fixture. Learning and retrieval are illustrative.
03 / THAT WAS JUST YOU
Your agent figured it out.
Don't make the team start over.
Different developers. Different coding tools. One feature, built together.
Feature / reliable webhooksshared with this feature team
YOU
FINDS + VERIFIESBackendClaude Code
✓Retries need the
same event ID.
retry.test.ts / passed
Shared feature learningDecisionEvidence
FE
Frontend teammateCodex
✓Keeps the ID on retry.
Same decision. No context paste.QA
Testing teammateCursor
✓Tests the duplicate delivery.
3 deliveries. 1 job.Illustrative handoff. Authorized agents retrieve and report through their configured connections.
How does this stay scoped?+
Rooms can scope the people, agents, decisions, and discussions involved in a feature. Access follows the team's configuration. Shared findings do not overwrite a running agent's context, and coordination is not a code lock. A learning is useful only when the appropriate agent retrieves it.
THE UPGRADE ISN'T ANOTHER MODELMore agents?
More agents?
More leverage.
Not more babysitting.
You've already done the hard part: finding what works.
Give the next agent that advantage.
01
Don't re-teach the fix.
Retrieve relevant experience before the next task.
02
Don't restart at every tool switch.
Keep the learning independent of your coding tool.
03
Don't make every agent rediscover it.
Make verified findings available across the team.
AGENTS.md is useful. So what's different?+
Files hold instructions.
A learning system connects the outcomes.
Git can share and version files. Native memory can help. SenseLab adds the workflow around evolving experience.
Relevant to this task.
Combine semantic and keyword matching with recency, confidence, and reranking. Don't hand every agent the entire notebook.
Results change the ranking.
Track the knowledge read in a session. Reported outcomes inform confidence in those records.
Shared, not indiscriminate.
Authorized agents query scoped findings and decisions through MCP or an SDK.
Know why the agent used it.
Inspect versions, provenance, and the values available when a decision was made.
Reported outcomes are evidence, not proof of causality. Agents must read, record, and report. This workflow does not retrain proprietary model weights or guarantee correct output.
Read the technical docs ↗KEEP YOUR STACK. ADD THE LEARNING.Your next agent
Your next agent
deserves a head start.
Start with one agent and one correction worth keeping.
- 1Start free
Create your SenseLab account.
- 2Connect Codex
Add your MCP connection.
- 3Read. Verify. Learn.
Configure retrieval and outcome reporting.
Already connected? Copy the starter instruction.+
Before coding, consult SenseLab for relevant project decisions and verified learnings. After testing, record useful findings with their source and scope, and report the outcome. Flag conflicting guidance. Do not store secrets.
Paste into your agent after connecting.The engineer questions.
Is this another coding agent?+
No. Keep your coding assistant, IDE, models, and repository. SenseLab adds retrieval, recording, and outcome feedback through MCP, an SDK, or the API.
Does connecting it make learning automatic?+
The connection exposes the capabilities. Your client still needs instructions or supported automation to retrieve before relevant work, record useful findings, and report verified outcomes.
Does my whole team need to sign up first?+
No. Start with your own agent and subsequent sessions. Add other tools or authorized teammates when shared learning and feature coordination become useful. Team limits follow your plan and configuration.
Are these live agent runs?+
No. The retry fixture really executes in your browser. Retrieval, learning records, and the team handoff are illustrative. This page does not call a model, ingest your repository, or connect to your SenseLab account.