Learning Loop AgentPack

Self-improving agent learning loop for Eve Horizon. Agents build persistent memory, discover reusable skills, and automatically fix platform friction.

Install

Add to your .eve/manifest.yaml:

x-eve:
  packs:
    - source: github:eve-horizon/learning-loop-agentpack
      ref: main

Sync the pack:

eve project sync

What's Included

Component	Type	Description
session-reviewer	Agent	Reviews completed sessions for learnings and friction
platform-improver	Agent	Fixes friction by opening PRs and issues
learning-brain	Skill	Composable skill — teaches any agent to learn
memory-manager	Skill	Memory CRUD procedures via Eve CLI
session-reviewer	Skill	Post-session review procedure
platform-improver	Skill	Friction-fixing procedure
review-cheap	Profile	Sonnet with low reasoning (cost-efficient review)
improver-capable	Profile	Sonnet with high reasoning (code-quality fixes)
post-session-review	Workflow	Triggered on job.attempt.completed
skill-review-heartbeat	Workflow	Every 6 hours — skill curation
batch-improvements	Workflow	Monday mornings — batch low-severity fixes

Opt-In Levels

Level 1: Passive Learning

Add the pack. Configure context.memory on your agents:

agents:
  my_agent:
    slug: my-agent
    skill: my-skill
    context:
      memory:
        agent: my-agent
        categories: [learnings, decisions, runbooks, context, conventions]
        max_items: 10
        max_age: 30d

Your agents start building memory. The reviewer extracts insights after each session.

Level 2: Active Skill Building

Add skill discovery:

    context:
      docs:
        - path: /agents/my-agent/skills/
          recursive: true

Agents now load and create reusable procedure docs.

Level 3: Repo-Committed Knowledge

Enable git access for skill promotion:

    policies:
      git:
        commit: auto
        push: on_success

Proven skills graduate from org docs to the project repo.

Customize

Override harness profiles in your manifest:

x-eve:
  agents:
    profiles:
      review-cheap:
        - harness: mclaude
          model: haiku
          reasoning_effort: low
      improver-capable:
        - harness: mclaude
          model: opus
          reasoning_effort: high

How It Works

Agent Session Completes
        |
        v
Session Reviewer (cheap, async)
  |-- Inward: Extract learnings -> Memory + Skills
  \-- Outward: Identify friction -> Structured findings
        |
        v (if severity >= medium)
Platform Improver (capable, async)
  |-- Fixable -> Open PR with evidence
  |-- Non-obvious -> Create GitHub issue
  \-- Skill fix -> Patch org doc
        |
        v
Slack Summary

Requirements

• Eve Horizon platform with job.attempt.completed event support
• Agent memory API (eve memory CLI commands)
• Org docs API (eve docs CLI commands)

License

MIT