Add mobius improve: self-diagnosing improvement loop

[AaronGoldsmith]

Summary

Mobius can evolve agent prompts via mobius evolve, but it never asks "what kind of agent am I missing?" or "what capability gap keeps causing failures?" This issue proposes a meta-layer (mobius improve) that diagnoses system-level gaps and proposes structural improvements.

Current State

The existing loop is: compete → judge → elo → evolve prompts → repeat

Everything improves within the existing agent pool. The system never:

• Identifies missing agent types or specializations
• Detects capability gaps from repeated failures
• Retires stale/underperforming agents
• Tracks why the system changed over time

Proposed: mobius improve

Three phases:

Phase 1: Diagnose (analyst.py + DiagnosisReport)

Analyze recent match history for patterns:

• Repeated failures — tasks where all agents score low → capability gap
• Narrow wins — barely-edged-out tasks → weak coverage
• Missing specializations — cluster tasks by topic, find uncovered areas
• Stale agents — high match count, declining win rate

Phase 2: Propose (ImprovementProposal model)

Generate typed proposals:

• create_agent — new agent for uncovered specialization
• retire_agent — remove underperformers
• split_agent — break generic agent into specialists
• add_capability — equip agents with new tools (e.g., WebSearch)
• system_change — structural improvements

Phase 3: Act (three modes)

Mode	Behavior
--dry-run	Print proposals only
--suggest	Create tracked proposals, human approves
--auto	Execute proposals directly

Agent Factory Pattern

Instead of hardcoding improvement logic, make it a competition:

mobius improve "We keep failing at web research tasks"
  → Spawns "architect" agents
  → Each proposes a different solution
  → Judge picks best proposal
  → Winner's proposal gets executed

Architect agents themselves evolve — they get better at proposing improvements because their proposals are judged on subsequent match outcomes.

New DB Table: proposals

proposals (
  id TEXT PRIMARY KEY,
  type TEXT,           -- create_agent, retire, split, system_change
  description TEXT,
  proposed_by TEXT,    -- agent_id or 'system'
  status TEXT,         -- pending, accepted, rejected, implemented
  evidence TEXT,       -- match IDs that motivated this
  outcome TEXT,        -- what happened after implementation
  created_at, resolved_at
)

Implementation Plan

1. src/mobius/analyst.py — match history analysis, DiagnosisReport model
2. ImprovementProposal pydantic model in models.py
3. proposals table in db.py
4. mobius improve CLI command
5. /mobius-improve skill (free Opus-powered version)

Related Ideas

• Agents that can create other agents and equip them with custom skills (stored in DB, not repo)
• Benchmark self-review that feeds back into the improvement loop
• Meta-learning: track which improvement strategies actually helped

🤖 Generated with Claude Code