Self-improving holon implementing Plan→Execute→Assess→Compound workflow loops with explicit knowledge crystallization. Generalizes compound engineering beyond coding to any domain where work produces learnable signals: writing, research, learning, problem-solving, design. Use when: (1) task quality matters and can improve over time, (2) patterns emerge from repeated work, (3) institutional knowledge should accumulate, (4) future tasks should benefit from past learnings, (5) "compound", "learn from", "improve process", "capture learnings", or multi-step workflows with review cycles. Orchestrates parallel research agents, multi-lens assessment, and structured knowledge codification into searchable documentation with YAML frontmatter. Implements λο.τ → λ(ο,Κ).τ where Κ is accumulated knowledge that compounds interest-like over time.
Core transformation: Task(ο) + Knowledge(Κ) → Output(τ) + Knowledge'(Κ') where Κ' ⊃ Κ (knowledge strictly grows)
Traditional workflows treat each task in isolation. Compound learning treats every task as a learning opportunity that improves future performance. Like compound interest, small improvements accumulate exponentially: each unit of work makes subsequent units easier, faster, and higher-quality.
Interest: A(t) = P(1 + r)^t
Learning: Κ(t) = Κ₀ × Σᵢ(1 + εᵢ) where εᵢ = learning from task i
┌──────────────────────────────────────────────────────────────────────┐
│ COMPOUND LEARNING LOOP │
│ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌──────────┐ │
│ │ PLAN │ ──▶ │ EXECUTE │ ──▶ │ ASSESS │ ──▶ │ COMPOUND │ │
│ │ (80%) │ │ (10%) │ │ (5%) │ │ (5%) │ │
│ └────┬────┘ └────┬────┘ └────┬────┘ └─────┬────┘ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌──────────┐ │
│ │Research │ │ Work │ │ Review │ │ Document │ │
│ │ Agents │ │ Agent │ │ Agents │ │ Agent │ │
│ │ (4+) │ │ (1) │ │ (5-12) │ │ (7) │ │
│ └────┬────┘ └────┬────┘ └────┬────┘ └─────┬────┘ │
│ │ │ │ │ │
│ ▼ ▼ ▼ ▼ │
│ plans/*.md outputs/ reports/ docs/solutions/ │
│ [category]/*.md │
│ │ │
│ ▼ │
│ ┌──────────────────────────────────────┐ │
│ │ KNOWLEDGE BASE (Κ) │ │
│ │ Feeds back into PLAN phase │ │
│ │ Making future work faster/better │ │
│ └──────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ (next iteration) │
└──────────────────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────┐
│ LAYER 1: COMMANDS (Workflow Orchestrators) │
│ /plan /execute /assess /compound /full-cycle │
│ │
├─────────────────────────────────────────────────────────────────┤
│ LAYER 2: AGENTS (Specialized Workers) │
│ Research: context-analyzer, domain-researcher, history-analyzer │
│ Execute: task-executor, progress-tracker │
│ Assess: quality-reviewer, pattern-detector, gap-identifier │
│ Compound: solution-extractor, prevention-strategist, doc-writer │
│ │
├─────────────────────────────────────────────────────────────────┤
│ LAYER 3: SKILLS (Knowledge Bases) │
│ Domain-specific expertise, schemas, patterns, procedures │
│ │
├─────────────────────────────────────────────────────────────────┤
│ LAYER 4: TOOLS (External Integrations) │
│ MCP servers, APIs, file systems, knowledge graphs │
└─────────────────────────────────────────────────────────────────┘
Purpose: Transform intent into structured, actionable plan using accumulated knowledge.
Time allocation: ~80% of total effort (most value comes from good planning)
Research agents (parallel):
| Agent | Function | Searches |
|---|---|---|
context-analyzer | Extract problem structure, constraints, goals | Current context |
domain-researcher | Gather domain standards, best practices | External sources |
history-analyzer | Find relevant past solutions, patterns | Κ knowledge base |
gap-analyzer | Identify unknowns, risks, dependencies | Plan vs. reality |
Output: plans/{task-title}.md with depth levels:
MINIMAL: # Simple tasks (<2 hours)
- Brief problem statement
- Acceptance criteria
- MVP approach
MORE: # Standard tasks (2-8 hours)
- Technical considerations
- Testing requirements
- Dependencies, risks
COMPREHENSIVE: # Major work (>8 hours)
- Architecture approach
- Implementation phases
- Alternative approaches considered
- Documentation plan
Κ integration: History-analyzer searches docs/solutions/**/*.md to find:
Purpose: Implement plan efficiently, tracking decisions and obstacles.
Time allocation: ~10% (quick when plan is good)
Work agent responsibilities:
Output: Task deliverables + execution metadata
class ExecutionLog:
decisions: List[Decision] # Choices made
blockers: List[Blocker] # Obstacles encountered
deviations: List[Deviation] # Plan changes needed
time_actual: Duration # Actual vs. estimated
Purpose: Multi-lens evaluation of output quality and process effectiveness.
Time allocation: ~5%
Review agents (parallel, domain-specific):
| Agent | Lens | Evaluates |
|---|---|---|
quality-reviewer | Output | Meets requirements? Robust? |
pattern-detector | Process | Reusable patterns emerged? |
gap-identifier | Coverage | Missing cases, edge conditions? |
risk-analyzer | Safety | Security, performance, reliability? |
efficiency-auditor | Process | Faster path existed? |
Output: Assessment report with actionable findings