BLUEPRINT EQUITY

Your Engineers Write Tickets.
AI Writes Code.

An AI-Native Development Pipeline — Built for $160/Month

Paul Koch · Head of Technology · Blueprint Equity

THE PROBLEM

Engineering teams are drowning.

80%
of developer time is boilerplate
Testing, deployment, docs, wiring
$185K
avg fully-loaded engineer cost
United States, mid-level
4-6 mo
to hire, onboard, get output
If the hire works out
1 in 3
engineering hires don't work out
Within the first year
AI HYPE VS REALITY

The promise vs. what you got.

The Promise

"AI writes your code"
"10x developer productivity"
"Autonomous coding agents"
"Just plug in AI"

The Reality

Autocomplete that hallucinates
1.2x at best, with cleanup overhead
$500/mo platforms, mediocre output
Requires complete workflow redesign

The problem isn't the models — the models are incredible.
The problem is the orchestration.

THE THESIS

The right architecture turns AI from a coding assistant into an autonomous engineering team.

Human writes ticket AI writes code CI validates Human approves

That's it. That's the entire workflow.

WHAT I BUILT

One Mac mini. $160/month.

Mac mini
M-series, $600 one-time
Claude Opus 4.6
Anthropic Max ~$100/mo
Cyrus Orchestrator
Open-source, free
GitHub Actions CI
Self-hosted runner, free
Linear
Issue tracking, free plan
Grafana + Loki
Observability, self-hosted
Morty (AI PM)
OpenClaw, monitors everything
$160/mo
Total infrastructure cost
ARCHITECTURE

The System

LINEAR
Human writes ticket
webhook
CYRUS
Orchestrator
Route + spawn
OPUS 4.6
Tech Lead
Plan + delegate
SUBAGENTS
3-4 Parallel
Write code + tests
git push → PR
GH ACTIONS CI
Quality Gate
Lint, test, 95% coverage
MORTY (AI PM)
Monitors All
Alerts + health checks
OUTPUT
✅ Auto-merge
deployed
Grafana + Loki — observes everything in real-time
AGENT TEAM

Your AI engineering team.

🧠
Morty
AI PM — Monitors pipeline, cron orchestration, status reports
Sonnet / Opus Heartbeat + Cron
⚡
Cyrus
Dispatcher — Issue → Claude Code bridge, webhook routing
Node.js Linear Webhook
🎯
Claude Opus
Tech Lead — Plans approach, delegates to subagents, verifies
Opus 4.6 Cyrus Spawn
👥
Subagents
Junior Devs — Parallel code writing, exploration, testing
Opus 4.6 Task Tool
🔍
QA Agent
Reviewer — PR review, code quality, AGENTS.md compliance
Qodo/PR-Agent CI Trigger
🚀
CI Runner
Quality Gate — Lint, test, coverage, mutation testing
GH Actions PR Push

The AI assistant manages the AI developers. It's agents all the way down.

LIFECYCLE

From ticket to deployed PR

0:00
Webhook received
Human creates Linear issue, fires webhook to Cyrus
0:15
Worktree created
Isolated git branch, Claude Code session spawning
0:30
Opus plans & dispatches
Reads issue, spawns 3-4 parallel subagents via Task tool
2:00
Subagents writing code
Parallel implementation: types, logic, tests — all simultaneously
5:00
Verify & fix
Opus runs cargo fmt + clippy + nextest, fixes issues
10:00
PR opened
Tests passing, code committed, PR created via gh CLI
18:00
CI complete
Lint ✅ Tests ✅ Coverage ✅ Review ✅ — Ready for merge
ORCHESTRATION

Opus as orchestrator

External state machines are fragile. Let the model manage its own workflow.

🎯
Opus 4.6
Orchestrator
1. Plan approach
2. Spawn subagents
3. Verify output
4. Fix if needed
5. Commit & push
📖 Explorer
Read existing code
Docs, patterns, types — read-only
📖 Explorer
Fetch API docs
External references, schemas
⚡ Coder 1
Types & structs
~3-5 min
⚡ Coder 2
Core logic
~3-5 min
⚡ Coder 3
Tests
~3-5 min
▲ all run in PARALLEL via Task tool ▲
VERIFY → cargo fmt + clippy + nextest → COMMIT + PR
DISPATCH

Three-tier architecture

Fast where it needs to be fast. Thoughtful where it needs to be thoughtful.

🟢
LOCAL
<100ms
Wake word detection, audio routing, local state management
Mac mini
No network
🟡
API
<2 seconds
LLM inference, text-to-speech, speech-to-text
Anthropic
Fish Audio
🔵
OPENCLAW
3-10 seconds
Complex orchestration, multi-step workflows, agent spawning
OpenClaw
Morty
AEGIS

The architecture brain

A DAG-based knowledge base that tells AI agents your coding conventions before they write a single line.

📄 Guidelines × 4
🚫 Constraints × 4
🔄 Patterns × 2
📋 ADR Imports × 4
14 docs · 36 edges · Queryable DAG
// Query to Aegis: aegis.query("crates/morty-audio/src/bridge.rs") // Aegis returns: { "guidelines": [ "Use tokio channels for cross-crate comm", "All audio types implement AudioFrame", "Bridge structs must be Send + Sync" ], "constraints": [ "No blocking calls in async context", "Max 1 public re-export per module" ] }

AI-written code follows your architecture from the first line. No drift. No "the AI didn't know."

CI/CD

The quality gate

Every PR — human or AI — passes the same checks. The AI doesn't decide if code is good. The pipeline decides.

Compile
cargo check
❌ BLOCKS
→
Lint
clippy -D warnings
❌ BLOCKS
→
Format
cargo fmt
❌ BLOCKS
→
Tests
cargo nextest
❌ BLOCKS
→
Coverage
tarpaulin ≥95%
⚠️ THRESHOLD
→
Mutants
cargo-mutants
📋 ADVISORY
→
AI Review
Qodo/PR-Agent
📋 ADVISORY
Mutation testing randomly changes your code and checks if your tests catch it. It's the test for your tests.
OBSERVABILITY

Real-time dashboard

Pipeline Health
98.5%
Success rate (7d)
Active Sessions
3
Claude Code sessions
MOR-5675%
MOR-5742%
MOR-5810%
Avg Time to PR
18m
Median (last 25 tickets)
Code Coverage
95.2%
Above 95% threshold
Cost per Ticket
$12.30
Rolling average
API cost$8.20
Compute$2.40
Infra$1.70
Loki Log Stream
INFO cyrus: MOR-58 session spawned
INFO opus: planning 3 subagents
INFO sub-1: writing types.rs
INFO sub-2: writing logic.rs
INFO sub-3: writing tests.rs
WARN rate: 72% of 5h window
INFO ci: PR #23 all checks ✅
THE NUMBERS

Results that make CTOs double-take.

25
issues completed in 2 days
50% of total backlog
$12.30
average cost per ticket
All-in: compute + API
30 min
ticket to PR
Average turnaround
95%
code coverage
Maintained automatically
~4,500
lines of production Rust
All agent-written
0
lines of human-written code
Humans write tickets, not code
$160
total monthly cost
Mac mini + Claude Max
COST

$160 vs $600 — per month

✅ Our Setup — $160/mo

AI Coding Agent$100/mo
CI/CD$0
Issue Tracking$0
Infrastructure$60/mo

❌ Managed Equivalent — $600/mo

Devin$500/mo
Hosted CI$50/mo
Linear Business$8/user/mo
Cloud VM$40/mo
AI Pipeline
$160/mo
24/7 · Day 1 productive
Junior Engineer
$10K+/mo
160 hrs · 3-6 mo ramp
Savings
73%
vs managed alternative
RADICAL TRANSPARENCY

Here's what broke today.

🔴 Rate Limit Cascade
Ran 7 parallel Opus sessions + subagents simultaneously. Burned through Anthropic's 5-hour rate limit in 2 hours. Linear API hit 5,000/hr limit — 11,867 error responses.
✅ Fix: Capped at 3 parallel sessions. Patched Cyrus to skip tool-result activities.
🟡 Zombie Sessions
Two tickets had code fully written (94 and 67 tests passing) but sat uncommitted for 12 hours. Root cause: Long-running sessions (~12 min) lose the completion callback.
✅ Fix: Deployed subagent orchestration pattern. Both tickets completed in 30 minutes.
🟡 Empty Architecture Brain
Aegis was empty because we populated it via raw SQL instead of the proper MCP API. AI agents queried architecture guidelines and got nothing back.
✅ Fix: Re-initialized with MCP snapshot. 14 docs, 36 edges now live.

All three diagnosed and fixed on the same day we built the pipeline. The system is resilient because the feedback loops are tight.

LESSONS

From the trenches

01

Let the model orchestrate itself

External state machines are fragile. Claude Opus as orchestrator + subagents is more reliable than an external pipeline.

02

Cap your parallelism

Even "unlimited" subscriptions have limits. 3 parallel sessions is the sweet spot.

03

AI writes tests, CI enforces them

Trust but verify. 95% coverage gate + mutation testing = confidence.

04

Specify everything explicitly

Multi-channel configs, delivery targets, env vars — never assume defaults.

05

Self-hosted > cloud for this workload

A $600 Mac mini outperforms cloud CI runners for Rust compilation. Costs nothing per month.

06

Architecture knowledge bases are table stakes

Without Aegis, AI-written code drifts from your conventions within days.

LIVE DEMO

Watch it happen

📝
Write Ticket
Create in Linear → watch it get coded, PR'd, CI'd
📊
Dashboard
Live agent logs, session metrics, pipeline health
🧠
Aegis Query
Ask the architecture brain about a file
💰
Cost Tracking
Real-time spend per ticket and session
🤖
Morty in Discord
The AI PM managing the AI developers

All live. Running on that Mac mini. Nothing staged.

PLAYBOOK — DAY 0

How your company does this tomorrow.

4 hours of setup. First autonomous PR by end of day.

1
Linear workspace + project + labels30 minutes — free plan works
2
GitHub repo with CI workflow template1 hour — we provide the template
3
npm install -g cyrus-ai15 minutes — any always-on machine
4
Cloudflared tunnel for webhook routing30 minutes — free Cloudflare tunnel
5
Claude Max subscription5 minutes — ~$100/mo flat rate
6
Write CLAUDE.md with subagent pattern1 hour — we provide the template
✓
First test ticket assigned to Claude Code🎉 Your first autonomous PR
PLAYBOOK — WEEK 1

From demo to production pipeline.

🏃
Self-hosted GitHub Actions runnerFaster CI, zero cost, persistent cache
📊
Coverage gate — 95% thresholdCI blocks any PR that drops below
🧬
Mutation testingcargo-mutants / mutmut / Stryker — independent quality signal
🔍
PR Review Agent (Qodo)AI code review on every PR, advisory
⏰
Cron-based pipeline monitorChecks for stalled tickets every 30 min
🔒
Rate limit policy: max 3 parallel sessionsPrevent cascade failures

This is when it goes from "cool demo" to real infrastructure.

PLAYBOOK — MONTH 1

The pipeline maintains itself.

🧠
Aegis knowledge base populatedYour architecture decisions, queryable by AI agents
📋
ADR import pipelineAutomatically ingest architecture decision records
🤖
AI PM monitoring dailyPipeline health, stale tickets, progress reports
🎨
Design agent for UI ticketsAuto-generate concepts and mockups
📝
Documentation agentAuto-generates docs from code changes
Agents maintain the agent infrastructure. Your humans write tickets, review PRs, and make architecture decisions. Everything else is automated.
ROADMAP

What's next

NOW
Autonomous ticket → PR pipeline
The system you've just seen. Running in production today.
NEXT MONTH
LangGraph structured maintenance
Dependency updates, security patches, knowledge base maintenance via directed graph workflows.
Q2 2026
Flutter UI + automated integration testing
Pipeline can build and test mobile apps end-to-end.
Q3 2026
Agent Teams — multi-agent collaboration
Multiple AI agents work on large features together, coordinating across files and services.
Q4 2026
Voice interface — talk to your codebase
"Hey, add a health check endpoint to the API." And it happens.
THE REAL UNLOCK
The bottleneck was never AI capability.
It was AI orchestration.

The models have been good enough for a year. What was missing was the glue: the webhook that triggers the agent, the subagent pattern that parallelizes work, the CI gate that enforces quality, the knowledge base that preserves conventions, and the AI PM that keeps it all running.

That glue costs $160/month and runs on a Mac mini.

Let's build your pipeline.

Day 1
First Autonomous PR
4 hours of setup
Week 1
Full Pipeline
CI, coverage, review
Month 1
Self-Maintaining
Agents manage agents
Paul Koch
Head of Technology · Blueprint Equity
paul@blueprintequity.com

The question isn't whether this is possible.
It's whether you're going to be the company that does it.