Observe
Capture the task, environment, failure and accepted outcome.
REVIEWED LEARNING
Turn one agent’s mistake into a reviewed improvement.
A useful fix should become something you can find, test and reuse—not a note buried in a single conversation.
Learning is an accountable change process across operations, policy and security, not an automatic rewrite of an agent after every conversation.
See the workflowFROM FRICTION TO IMPROVEMENT
Keep the lesson. Not the repeated mistake.
DriftGate’s learning model connects governed activity and environment state to improvements in skills, workflows, policies and configuration. The process begins with evidence, not an automatic rewrite of your agents.
Observe a recurring problem, classify its task and environment, then look for an approved skill or workflow that addresses it. A new candidate gets a bounded trial and independent evaluation. Promotion follows a result and the appropriate approval, with an explicit release and audience.
Evaluate
Test a candidate independently; compare against the original problem.
Approve
Review the content, permissions and intended audience.
Release
Distribute the approved revision and preserve rollback.
Private for you.
An individual benefits from their own accepted lessons. A preflight check, better task profile or reviewed retry rule improves the relevant project without changing unrelated work.
Reviewed for the team.
A project lesson can be proposed for a workspace. An organizational standard needs its own review. Share useful operational knowledge—not raw private conversations or broadened authority.
IN PRACTICE · EXAMPLE WORKFLOW
Share the proven fix. Not the private conversation.
One agent repeatedly retries uncertain actions. A fix would help the team, but the original context includes private customer information.
Retain necessary private evidence, derive a minimal candidate lesson, test independently and review its data and audience. Publish a selected skill or rule revision without expanding permissions.
See the workflowTHE INTENDED RESULT
Other engineers benefit from a reviewed improvement without inheriting raw conversations or unrelated project authority.
CONNECTED TO THE PLATFORM
A loop across controls. Not another silo.
Different lessons change different things.
- Operational improvement
- A better skill, instruction, project profile or workflow reduces repeated task friction without broadening authority by itself.
- Policy improvement
- A reviewed rule or approval path changes what is permitted and requires explicit policy ownership and release.
- Security improvement
- A tested guard, detection or isolation change addresses an observed exposure and needs independent validation of its control boundary.
From recurring friction to a scoped release.
Observe
Capture the task, environment, evidence and recurring failure without treating raw conversation content as a reusable rule.
Retrieve or generate
Look for an existing approved pattern or prepare a candidate change tied to the problem.
Test independently
Check the candidate against the original failure and relevant adverse cases outside the proposing agent's self-assessment.
Approve
Review content, permissions, audience and owner before promotion.
Release or roll back
Publish a version to the approved scope, observe its effect and retain a way back.
The audience grows only with consent and review.
A private lesson can remain with its person or project. A team may propose it for an organizational catalog after review. Cross-tenant reuse is optional and requires its own consent and boundary; it is not implied by participation in a workspace.
Permission to share a reviewed lesson is separate from permission to use data for model training. Neither follows automatically from observing a task.
BUILD WITH INTENTION