Company profile · 2 min read
SenseLab: continual learning from agent outcomes
SenseLab describes a shared learning layer where agents record what they learned, read briefings and update confidence from outcomes.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
SenseLab gives repeated agent runs a shared memory and outcome layer so findings can be reinforced or degraded over time.
What it does
The public documentation describes recorded runs, shared rooms, confidence scores and outcomes that strengthen or decay findings. The homepage says the layer works across agent frameworks without requiring a framework change.
Buyer and task
Teams running repeated agent tasks that want knowledge and judgments to compound across sessions.
Workflow boundaries
The checked public pages do not show an independent improvement measurement. The 1M tasks and 10,000 teams displayed on the homepage are company claims.
What I would ask
Ask how bad outcomes reduce confidence, who can edit shared knowledge, and how sensitive traces are isolated between teams.
Why it fits
SenseLab’s public explanation starts with a real limitation of agent systems: a session ends, the work disappears and the next run starts with the same model. Its proposed loop records what was asked, read, decided and done; stores findings in shared rooms; then updates confidence from real outcomes. The feature page describes confidence scores that rise when a finding leads to a good decision and decay when it does not, plus cross-agent reads and queries by confidence threshold.
That is a useful distinction from a memory store that only retrieves whatever was written last. A team running repeated support, research or operations tasks could use the records to compare what an agent believed with what happened next, then decide which findings deserve to be reused. The public pages also say the system is framework-agnostic and may train a private model from accumulated decisions. Those are product claims; the checked public pages do not show an independent improvement measurement or a live integration.
The first test should be a repeated task with an explicit outcome and a deliberately wrong finding. Ask how the wrong result degrades, who can edit shared knowledge, what a reviewer sees and how data from different teams is isolated. The homepage’s “1M+ tasks” and “10,000+ teams” are company claims, not adoption evidence established here.
Quick facts
| Field | Sourced detail |
|---|---|
| Buyer | Teams running repeated agent tasks |
| Core loop | Record runs → share findings → validate outcomes → update confidence |
| Integrations | Framework-agnostic positioning; docs describe shared rooms and agent reads |
| Evidence limit | No independent improvement or live-integration result documented |
Sources checked
official Speedrun profile · company homepage · source page · source page
Sources checked — September 20, 2026.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗