Mythical Man Month

SKILL: The Mythical Man-Month

Strategic framework for managing complex system development, revealing why intuitive approaches to scaling, coordination, and change management systematically fail.

DECISION POINTS

Scope Change Response Matrix

Team Scaling Decision Tree

IF task has zero interdependencies
  THEN people scale linearly (rare in software)

IF task requires coordination between workers
  ├─ IF coordination is <20% of effort per person
  │   THEN add people cautiously (2-3x max team size)
  ├─ IF coordination is 20-50% of effort
  │   THEN keep team small (3-7 people max)
  └─ IF coordination is >50% of effort
      THEN redesign task to reduce dependencies

IF project is late AND <50% complete
  THEN extend deadline OR cut scope (never add people)

IF project is late AND >50% complete
  THEN extend deadline only (adding people guarantees failure)

Agent Architecture Coordination Strategy

IF agents need shared state
  ├─ IF state changes <10 times/minute
  │   THEN use single coordinator with state ownership
  └─ IF state changes >10 times/minute
      THEN redesign to eliminate shared state

IF agents need to negotiate outcomes
  ├─ IF <3 agents total
  │   THEN allow peer-to-peer communication
  └─ IF ≥3 agents
      THEN use mediator pattern (single orchestrator)

IF task requires conceptual unity
  THEN single orchestrator with specialized tools
  NOT peer-to-peer agent democracy

Documentation Priority Decisions

IF you can't write clear spec in 2 pages
  THEN design isn't ready for implementation

IF writing documentation feels painful
  THEN you haven't decided what you're building

IF spec and code diverge
  ├─ IF spec is more accurate → fix code
  ├─ IF code is more accurate → update spec
  └─ IF both are wrong → stop and redesign

Priority order:
1. User interface/API spec (what system does)
2. Architecture overview (how pieces fit)  
3. Individual component specs
4. Implementation details (last)

FAILURE MODES

The Late Project Death Spiral

Symptoms: Project behind schedule, management proposes adding people, velocity decreases further Detection: "We're 3 months late, let's hire 5 more developers" Root Cause: Training overhead + repartitioning costs exceed productivity gains Fix: Accept delay OR cut scope OR redesign for different partitioning. Never add people to late project in final stages.

Committee Architecture Syndrome

Symptoms: System has multiple interaction paradigms, inconsistent interfaces, users can't form mental model Detection: Every feature requires learning new interaction pattern Root Cause: Democratic design process without single architectural vision Fix: Designate single architect for external interfaces. Implementation can be distributed, architecture must be unified.

The 90% Done Trap

Symptoms: Progress reports show 90% complete for months, testing keeps finding integration issues Detection: No concrete deliverable milestones, "almost done" status persists >2 weeks
Root Cause: Underestimated integration complexity, no sharp completion criteria Fix: Define binary milestones (passes all tests, user can complete workflow X). Allocate 50% of schedule to integration/testing.

Requirements Stability Delusion

Symptoms: Every change request treated as scope creep emergency, architecture can't accommodate variations Detection: "But that's not what we originally agreed to!" repeated frequently Root Cause: Assuming requirements won't change vs. designing for inevitable change Fix: Plan for 25% requirement changes. Build clean interfaces. Treat change as normal, not exceptional.

Second System Effect

Symptoms: Reimplementation includes every withheld feature plus kitchen sink, loses original system's elegance Detection: New version has 3x features but takes 5x longer to learn Root Cause: Architect compensating for first system's limitations by including everything Fix: Maintain discipline. Second system should be refined, not inflated. Test conceptual integrity with naive users.

WORKED EXAMPLES

Multi-Agent Code Review System

Scenario: Building automated code review with 4 specialized agents (syntax, security, performance, architecture).

Naive Approach: Peer-to-peer agent communication where each agent reviews independently then they negotiate final score.

Brooks Analysis Applied:

1. Communication overhead: 4 agents = 6 communication links. Each agent must understand others' reasoning.
2. Conceptual integrity: Who owns the final "accept/reject" decision? Committee vote creates inconsistent standards.
3. Change impact: New review criteria requires updating all 4 agents' negotiation logic.

Expert Decision Process:

• Architecture choice: Single orchestrator agent with 4 specialized tool agents
• Reasoning: Orchestrator maintains consistent review philosophy, tools provide analysis
• Interface design: Each tool returns structured data + confidence score, orchestrator decides
• Change management: New criteria only requires updating orchestrator logic

What novice misses: Tries to make agents "smart" and autonomous What expert catches: Minimizes coordination points, preserves decision consistency

Result: 5x simpler debugging, consistent review standards, easy to modify review criteria

Late AI Training Pipeline Project

Situation: ML training pipeline 2 months behind, manager wants to add 3 GPU engineers.

Decision Process:

1. Task analysis: Pipeline requires sequential stages (data → preprocessing → training → validation)
2. Current bottleneck: Data preprocessing taking 3x expected time
3. Communication overhead: New engineers need 2-4 weeks to understand custom data formats
4. Training burden: Existing 2 engineers must stop work to onboard newcomers

Brooks-informed decision:

• Don't add people: Training cost (8 person-weeks) exceeds remaining schedule slack (6 weeks)
• Cut scope instead: Defer advanced preprocessing features, use simpler baseline
• Extend deadline: 4 weeks for current team to finish vs. 8+ weeks with added people

Outcome: Delivered in 10 weeks vs. estimated 12+ weeks if people added

QUALITY GATES

Design Phase Completion

• Single architect can explain system in 2-page overview
• External interfaces documented before any implementation starts
• User mental model tested with 3+ naive users
• All major components identified with <5 interfaces between them
• Change scenarios (25% requirement shift) don't break architecture

Development Phase Milestones

• Each milestone produces working system (not just components)
• Integration budget allocated (25% of total effort minimum)
• Communication overhead measured: coordination time <20% per developer
• Team size stable: no additions after 50% schedule point
• Documentation current: spec changes drive code changes, not reverse

Pre-Ship Quality Gates

• Conceptual integrity verified: naive user can predict system behavior
• All planned features implemented OR explicitly descoped with stakeholder sign-off
• System testing complete: full end-to-end workflows pass
• Performance meets requirements under realistic load
• Documentation sufficient for maintenance team handoff

NOT-FOR BOUNDARIES

This skill should NOT be used for:

• Perfectly partitionable tasks (data processing, batch jobs) → Use parallel processing patterns
• Research/exploration projects → Use rapid prototyping methodologies
• Small teams (<4 people) → Coordination overhead is minimal, other factors dominate
• Well-understood problems with stable requirements → Use standard SDLC approaches
• Real-time systems with hard latency constraints → Use real-time engineering principles

Delegate instead:

• For technical architecture decisions → Use system-design skill
• For individual productivity → Use deep-work and focus-management skills
• For requirements gathering → Use stakeholder-management skill
• For team dynamics → Use team-leadership skills
• For project tracking → Use agile-project-management skill

Key boundary: Brooks's insights apply when coordination overhead becomes significant factor. If your team spends <10% time on coordination, other management approaches may be more relevant.