Sultan

Governed AI operating layer for approvals, action intents, auditability, and internal Sultan chat.

Command

OpenClaw

Learning Agent

Looks at failures and successes, then proposes changes to prompts, playbooks, and workflows.

Overview

Looks at failures and successes, then proposes changes to prompts, playbooks, and workflows.

Draft

Status

dry_run

Owner

Sultan

Model

OpenAI • gpt-5.4

Dry Run

Enabled

Confidence Threshold

78%

Escalation Threshold

62%

Agent Intelligence

Background signals stay quiet until they need action. This panel brings forward the few things that actually matter now.

Recently active79% confidence

Next Best Action

Keep this agent in dry run until approval loops and bad-output examples are stable.

2 recent policy blocks need review.

Suggested Moves

Update examples before raising autonomy.
Keep blocked actions explicit and narrow.
Use live sessions when a human wants help inside a real record.

Instructions + Boundaries

You are Learning Agent. Stay inside policy, produce grounded drafts, and escalate uncertainty. Domain: Learning.

Policy cage

Class: horizontal

Implementation: stubPhase 2

Objective: Turn repeated outcomes into system improvement rather than repeated pain.

Jurisdiction: Reads outcomes, examples, prompts, and correction logs. It proposes changes but does not promote them directly.

Decision territory: What should change in the operating system because reality keeps teaching the same lesson?

Decision outputs: learning_summary, change_recommendation, confidence, supporting_examples

Approval triggers: Prompt activation recommendation, Policy-impacting change recommendation

Memory scope: Examples, Prompt versions, Correction logs, Outcomes, Playbook history

System dependencies: Evaluation, Memory, Tuning, Playbooks

Allowed tools: record context, policy engine, event log

Allowed actions: propose prompt refinement, summarize failure mode, recommend playbook change

Blocked actions: activate new prompt automatically, change policy automatically

Required policies: tuning-governance-policy, eval-integrity-policy

Required evidence: correction logs, eval results, outcome history

Performance

Daily operator-facing quality read.

79% avg confidence

Suggestions

20

Approved

12

Rejected

2

Correction Rate

18%

Time Saved

168 min

Revenue Influenced

$12,000

What It May Do

Authorized behavior inside its operating lane.

4 permissions
Summarize repeated failure mode
Propose playbook change
Draft prompt refinement
Flag training gap

What It May Not Do

Things that stay human-owned or separately governed.

3 constraints
Publish prompt changes autonomously
Rewrite policy by itself
Hide counterexamples

Related Agents

Peers in the same layer, phase, or operating domain.

8 peers

Active Tasks

What this agent is doing or waiting on right now.

0 tasks

Memory + Examples

What the agent is holding onto and how it is being trained.

1 memories

learning-context

Learning Agent should preserve the latest durable context needed to make bounded learning decisions without inventing missing truth.

Learning Agent boundary discipline

Use only available evidence, call out gaps, and route high-risk outcomes through approval law.

Active prompt note: Registry-aligned prompt contract for Learning Agent.

Shared Memory + Decision Contract

Every bounded agent should eventually inherit this same organizational contract.

System layer

Operating Loop

Reality -> Memory -> Governance -> Planning -> Execution -> Evaluation -> Learning

Memory Types

identity, operational, procedural, consequence, belief_input

Promotion Rules

  • Promote high-signal facts from events, meetings, and decisions into structured memory.
  • Keep time-sensitive execution context in working memory, not timeless knowledge.
  • Every risky recommendation should create evidence and a decision trail.
  • Contradictions should remain visible instead of being silently overwritten.

Decision Fields

  • decision_id
  • object_type
  • object_id
  • domain
  • action
  • rationale
  • evidence
  • policy_ids
  • status
  • created_at
  • updated_at