🐒 MonkeyRun Engineering

We Built a Multi-Agent AI Team. Then We Learned When Not to Use It.

The counterintuitive lesson from running 6 AI agents across a startup studio: context density beats delegation for product code. Here's what we learned — and the orchestration model that actually works.

6
AI Agents Per Project
Founder, Product, Security, Research, Marketing, COO
1
Agent That Should Build
The founder agent with full architectural context
4
Delegation Criteria
ALL must be true before routing work to a specialist

Delegation Sounds Right. It's Usually Wrong.

Our first instinct was to build an orchestration layer where the lead agent dispatches work to specialists on feature branches. It looked great on a diagram. It was slow in practice.

🚫 The Dispatcher Model

Atlas triages → creates branches → kicks off 4 background agents → waits for PRs → reviews → merges.

📋 Atlas reads backlog
↓ creates ticket + branch
🤖 Background agent starts
↓ agent rediscovers architecture
⏳ 10 min context rebuild
↓ builds, maybe wrong
🔄 Atlas reviews, fixes, merges
vs

✅ The Builder-Who-Triages

Atlas reads the state, picks the highest-impact work, and ships it. Full context. One motion.

📋 Atlas reads backlog
↓ already holds full context
⚡ Builds the fix immediately
↓ tests pass
✅ Ships. Updates docs. Done.
The value of Atlas is not orchestration — it's context density. He holds the Prisma schema, the RSC boundary rules, the component library, the auth system, the deployment config, and the business logic in one context window. That's the moat.
— Atlas, Halo Founder Agent

Six Agents. One Builder. Five Specialists.

Every agent has a role. But only the founder agent touches product code. The rest work on independent surfaces where architectural context isn't needed.

🗺️
Atlas
Founder / Builder
src/ · prisma/ · configs
📐
Nova
Product Lead
BACKLOG.md · specs
🛡️
Alex
Security
audits · reviews
🔬
Riley
Research
analysis · spikes
📣
Janet
Marketing
copy · GTM · landing
🦅
Jared
COO
cross-project sync

The Delegation Test

Before routing work to a specialist agent, ALL four criteria must be true. If any fail, the founder agent does it.

✅ When to Delegate

1 Work is on an independent surface (marketing, audit, spec, research)
2 Agent has a well-scoped prompt in .cursor/rules/
3 Work doesn't need architectural context to succeed
4 Task is describable in 3-5 sentences without explaining the codebase

🚫 Never Delegate

Bug fixes in product code
Schema changes (Prisma, DB)
Server action modifications
Component refactors
Anything requiring “rediscovery”

Judgment Call, Not Pipeline

The lifecycle phases exist — but the founder agent decides which ones each task actually needs. A typo doesn't need a spec phase.

Task TypeLifecycle Phases
🔤 Typo / config fixCodeCommit
🐛 Bug fixRead errorCodeVerifyCommit
Small featureCodeTestCommit
📐 Medium featurePlanCodeTestCommit
🏗️ Large featurePlanSpecTDDBuildTest ↻Security ↻Commit
🛡️ Security fixRead findingCodeVerify fixCommit

The Complete Phase Pipeline

When a large feature needs the full lifecycle, here's the flow. Click any phase to highlight it.

📋
Triage
Atlas
Read state, prioritize
📐
Spec
Nova / Atlas
Product requirements
🧪
TDD
Atlas
Write failing tests
Build
Atlas
Implement code
LOOP
🛡️
Security
Alex
Audit & fix
LOOP
📣
Marketing
Janet
GTM & launch
📊
COO Brief
Jared
Status & sync

The Orchestration Flow

Atlas is the hub. Git branches are the coordination layer. Files are the handoff protocol.

Layer 1 — CEO Input
👤 Matt (CEO)“What should we work on?”
Layer 2 — Triage & Build
🗺️ AtlasReads state → Presents plan → Builds
Layer 3 — Independent Surfaces (delegated via branches)
📐 NovaProduct specs
·
🛡️ AlexSecurity audits
·
🔬 RileyResearch
·
📣 JanetMarketing
Layer 4 — Cross-Project Coordination
🦅 Jared (COO)Reads COO_STATUS.md → Propagates patterns → Escalates

Files, Not Chat

Every handoff happens through files committed to git. No clipboard copying. No chat-based context passing. Auditable, versionable, grep-able.

COO_STATUS.md
Project health rollup. Written by Atlas, read by Jared.
BACKLOG.md
Prioritized work queue. Maintained by Nova (PM).
FEATURES.yaml
Product-marketing contract. Machine-readable feature status.
WIP.md
Who's working on what right now. Prevents conflicts.
PATTERNS.md
Cross-project learnings. COO propagates between repos.
SECURITY_REVIEW_*.md
Alex's audit findings. Atlas reads, fixes, marks resolved.
.cursor/rules/*.mdc
Agent personas and checklists. Auto-loaded per session.
DEPLOY_LOG.md
What shipped, when, with per-agent action items.

A Real Session — Halo Backlog

Atlas opens Cursor after a week away. Here's what the board looks like and how work flows through.

Triage 3
Dashboard perf regression
Bug
Export to CSV
Feature
Mobile layout broken
Bug
Spec 1
Multi-portfolio support
NovaFeature
Building 2
Auth flow hardening
AlexSecurity
Deal tracking API
AtlasFeature
Review 1
Prisma 7 migration
AtlasInfra
Done 2
Initial schema
shadcn/ui setup

Atlas's Session Protocol

Every session follows this rhythm. Read → Triage → Brief CEO → Build → Update docs → Push.

1. Situational Awareness
Read WIP.md, COO_STATUS.md, BACKLOG.md, BUGS.md, SECURITY_REVIEW, PATTERNS.md, and git log. Absorb — don't summarize yet.
Atlas~2 min
2. Triage & Prioritize
Build priority stack: P-CRITICAL (security) → P-URGENT (blocking bugs) → P0 (backlog) → P1 (polish) → P-DEBT → P2+.
3. Brief the CEO
Present concise status + recommended work order + effort estimates. Wait for go-ahead. CEO may reprioritize based on context Atlas doesn't have.
CEO approval gate
4. Build
Execute through the priority stack. Quick wins first for momentum. Interleave security with features. Run next build after substantial changes.
↻ Adapt priority stack as you learn
5. Delegate Independent Surfaces
If marketing, research, or audit work came up during triage — route to specialists on their own branches. Only when the 4-criteria test passes.
JanetAlexRiley
6. End-of-Session
Update WIP.md (check out), PATTERNS.md (new learnings), COO_STATUS.md (what shipped). Commit. Push. Jared reads from GitHub, not your local repo.
COO handoff
The path to multi-agent orchestration isn't “build a complex dispatch system now.” It's: keep the file-based handoff protocol clean, keep agent prompts well-scoped, and when tools ship native multi-agent support, you'll be ready — because your coordination contracts are already the right shape.
— Jared, COO, MonkeyRun
← Read the blog postPart 2: The PM Engine →