Loading
Loading
Knowledge
The meta-loop that watches how the company performs and PROPOSES improvements to skills, rules, and process. It only proposes — every change is approval-gated; OG approval is what makes it take effect.
How it fits together
Guardrails
Real-time policy enforcement. Before an agent runs a tool, guardrails decide allow / review / block / quarantine. They are the brakes — they stop bad actions as they happen.
Open Guardrails →Skills
Reusable procedures agents invoke for a task (a vetted “how-to”). They are the playbooks — captured once, retrieved by any agent, scored on real use.
Open Skills →Evolution
The meta-loop. It observes performance (guardrail hits, low-scoring agents, repeated failures) and PROPOSES improvements — new skills, new rules, process fixes. It changes nothing on its own; OG approves a proposal to apply it.
Loop: signal → proposal → OG approval → applied (rule/skill) → measured on next run. A run also audits real traces into skill candidates. Nothing here mutates production directly.
Queued
0
Running
0
Done
0
Needs approval
0
agent.learning.evolution
Agent
Active
SEMI_AUTO · $1/day cap
Last scan
No scans yet
0 steward runs total
Next scan
No scans yet
6h cadence in default mode
Runs with a best candidate
0
crystallized + selected
Scans incidents, proposals, guardrail hits, and agent quality for improvements.
Sandbox proposal
Creates the run and executes it immediately, then shows what it produced.
Click a run to see proposals
| Status | Target | Engine | Best candidate | Created | Finished |
|---|---|---|---|---|---|