AI Is an Amplifier: Harness Engineering for 3x Velocity — GOTO Accelerate Chicago 2026

Accelerate Chicago — Craft software with confidence in 2026

// speaker & sponsor

Where I work

They pay my bills

rivet.work/careers

Phil Borel

Speaker

philborel.com

// detroit software developers

Detroit Software Developers

Monthly meeting
Software development focus
Technical Talks | Good conversation

detroitdevelopers.com

// intro

AI is an Amplifier

Harness Engineering for 3x Velocity

Phil Borel · GOTO Accelerate Chicago · June 2026

GOTO Accelerate Chicago talk page QR code

detroitdevelopers.com/talks/goto-accelerate-chicago

// act I — the goal

Act I · February 2026 · All-Hands

3x Product
Development Velocity

BHAG - Big Hairy Audacious Goal

by Q1 2027

// how we got started

The BHAG gave us an excuse to take Agentic Dev seriously

Learned - read papers & blog posts; watched talks; listened to podcasts
Convened - started a weekly AI Roundtable where devs share what's working & what's not
Experimented - every engineer using Claude Code; open-door policy on tools & workflows; no caps on skills, CLAUDE.md updates, etc
Planned - created quarterly goals tied to outcomes to track progress against the BHAG
High-volume learning environment, not a slow-roll adoption process

// what does 3x mean?

What does 3x even mean?

Vibes - execs, CS, sales, product feel like we shipped 3x the value
Metrics - DORA metrics (CFR, MTTR, DF), 3x PRs, bugs / time
Outcomes - 3x projects shipped; users get more value more quickly
All three matter. None are sufficient alone.
Risk: 10x some metrics while seeing modest gains on outcomes
Increase velocity | reduce bugs - Bugs / Release; PRs / Sprint; Projects / Quarter

// work in progress

A working case study.
Not a success story.
Yet.

Progress from Feb -> June:

~2x velocity increase vs last year
much of the acceleration came from process changes, code refactors & quality initiatives - not from specific tools or Agentic Dev
<1 incident per month down from >1 per 2-week sprint

// act I take-away

Act I take-away:
Set a Goal

Vibes · Metrics · Outcomes

// act II

Act II

AI is an
Amplifier

y = mx + b || Inputs -> Outputs

// the research

AI is an amplifier.

DORA 2025 (5,000 professionals): high-performing orgs see 50% fewer customer incidents. Struggling orgs: 2× more.
Carnegie Mellon 2026 (807 open source repos): 281% increase in lines added month one. Month two: velocity gone, back to baseline. Complexity +41% permanently.
In the wild (Amazon retail): spike in outages → senior sign-off mandate.
"AI is moving organizations in different directions." - Laura Tacho, AWS

// what AI amplifies

What does AI amplify?

Good architecture → maintainable code
Bad architecture → friction increases with every new feature
Good specs → accurate, testable implementations
Vague specs → confidently wrong code that passes tests you don't understand
Consistent patterns → idiomatic output on the first try
n divergent patterns → tech debt increases agentically

// the implication

Best practices matter more with AI, not less.

Agentic Dev has unlocked a renewed focus on quality

// keeping quality in check

4-Step Review Process

1. Architectural review - human, before code is written
2. Automated review - /review-pr skill, SonarCloud, AI code review (Copilot || Greptile), linting, tests (all in CI)
3. Human code review + manual QA - focused on the feature, based on test steps in the PR
4. Full-app UAT/regression - end of every sprint (working toward replacing with Playwright)

// where we are focused

Where we are focused

Code quality - refactors
DevOps - infra updates focused on reduced dev process
QA - focus on Playwright & test automation
All driven by metrics

// act II take-away

Act II take-away:
Focus on the Fundamentals

Leverage AI to improve your codebase

// act III

Act III

Agentic Dev
Maturity

Prompt · Context · Harness

Agentic coding maturity blog post QR code

detroitdevelopers.com/blog/agentic-coding-maturity

// the three layers

Each outer layer contains and depends on the inner ones

// layer 1: prompt engineering

Prompt Engineering

Telling an agent what to do effectively
Written prompts, rules, skills
In practice: almost entirely markdown optimization
Tight feedback loop - edit a file, run a session, see what changed
Ceiling: ~+1x velocity
systems to measure & optimize prompt effectiveness (Chip Huyen)

// layer 2: context engineering

Context Engineering

Managing what knowledge the agent has access to - and when
PILRs - Persistent Indexed Learning Repositories
Type 1 (Ephemeral): per-feature planning docs, test plans
Type 2 (Evergreen): architecture docs, system design, API contracts
Type 3 (Cumulative): solved problems, incident patterns, institutional memory
Ceiling: ~+3x velocity

// context engineering: claude.md

CLAUDE.md: map, not encyclopedia

Large files degrade agent performance due to "the dumb zone" (context window)
When everything is "important", agents fall back to local pattern-matching
~100 lines max - structured as a map with pointers to deeper sources of truth
Checked into the repo. Reviewed in PRs. Evolved alongside the codebase.
Team artifact, not personal config

// context engineering: pilrs

PILRs: knowledge bases that learn

Structured markdown files + an index.json that tells the agent where to look
The agent loads only what it needs - reducing context window bloat
Note: PILRs continue to evolve as the agent does the work
Every fix, every new pattern, every bit of domain knowledge the agent picks up → feed back to the PILR
Unlike a wiki that goes stale, a PILR is maintained by the thing that reads it

// context engineering: pilrs in practice

PILRs at RIVET - temp/projects directory structure with high-level planning docs, per-PR plans, and test plans — PILRs at RIVET

PILRs in the Superpowers plugin - docs/superpowers/plans and specs directories — PILRs in Superpowers

github.com/obra/superpowers QR code — PILRs in Superpowers

// layer 3: harness engineering

Harness Engineering

Programming around the model - systems that run agents for you
Multi-step autonomous workflows with human review of outcomes
Deterministic steps where reasoning isn't needed
Expands what work gets done at all - not just speed. Collapses a full ticket - research, implementation, PR - into a parallelizable workflow
Ceiling: ~+10x velocity (S-curve)
Every team has a long tail of valuable work that never makes the top of the backlog

// case study: odradek

Odradek - Bug Resolution Agent

Built on the Claude Code SDK - Mac desktop app with cloud-hosted database for multiplayer support
Investigates - reads source, checks git, consults PILR knowledge base, pulls context from Notion (MCP) & GitHub (CLI)
Fixes - scoped, surgical edits to the relevant files
Verifies - regression checks, test additions & CI runs
Opens a PR - agentic-review phase + human-review phase
One-shot resolution rate: 80%

// odradek: resolutions view

// odradek: impact on the backlog

Odradek impact on the backlog

P1s (ship-within-a-week): used to sacrifice one engineer per sprint on rotation. Odradek bought back ~half an engineer.
P2s (fix-within-a-quarter): used to stack for months. Now addressed as they come in.
P3s (nice-to-fix): were permanent backlog residents. Some are actually getting fixed now.
Work that would never have happened at the current team size

// the three ceilings

// act III take-away

Act III take-away: where to invest your tokens for optimal returns

Beginner: Start with prompt engineering. Short feedback loop, transferable skills, necessary foundation
Intermediate: Context engineering. Build the knowledge layer. Start with PILRs where agents are most confused
Advanced: Harness engineering. Pick a narrow, repetitive workflow. Measure the one-shot rate
The balance: increasing velocity - maintaining quality

// what about product & design?

Product Convergence: The new bottleneck

Prototyping
Spec writing
Feature work - TBD

// close

Next week

Set a goal: identify an outcome & turn it into a goal
Focus on fundamentals: leverage the goal to prioritize tech debt & quality initiatives
Invest in Agentic Dev maturity: focus on prompt, context & harness engineering in that order. Read the docs @ claude.md
Try Superpowers: github.com/obra/superpowers - it's easy!
None of this requires a BHAG or an executive mandate

Questions? Ask me in the app — App Store and Google Play QR codes

// appendix

Appendix

Reference
Material

Extra slides for Q&A

// RIVET - prompt engineering

Far along, near the ceiling

Deep investment in CLAUDE.md, rules, skills
Almost entirely markdown optimization
Tight feedback loop - edit, run, observe
Most of the returns are already captured

// RIVET - context engineering

Early-mid, significant upside

PILRs pattern is right; infrastructure isn't finished
Type 1 (Ephemeral) solid, Type 2 (Evergreen) in progress, Type 3 (Cumulative) nascent
Starting to invest in shared hosting & data infra

// RIVET - harness engineering

Very early, asymmetric upside

Odradek is our first real harness - 80% one-shot rate
S-curve: slow start, steepest returns ahead
This is where we're investing

// appendix: odradek roadmap

Odradek Roadmap

Event-driven triggers - auto-fire on bug ticket creation
Cloud-hosted dashboard - non-engineers see work in progress & resolution status
MS Teams integration - ask about bugs & initiate workflows from chat
Parallel issue processing - git worktrees for concurrent work
Ephemeral test environments - CS/product verify fixes themselves
Model routing - right model tier for each task; balances cost & effectiveness