// welcome
Detroit
Developers
monthly meetup 
detroitdevelopers.com QR code
detroitdevelopers.com
// thank you to our sponsor
RIVET Careers QR code
We're hiring →
rivet.work/careers
// your organizers
Phil Borel
Phil Borel
Organizer
Phil Borel QR code
philborel.com
detroitdevelopers.com QR code
detroitdevelopers.com
Louis Gelinas
Louis Gelinas
Organizer
Louis Gelinas QR code
linkedin.com/in/louis-gelinas
// intro
Agentic Software
Development
Advanced Practitioner's Guide 
Phil Borel · RIVET · Detroit Developers · April 2026
// intro
"A working case study.
Not a success story.
Yet."
// intro
What does 3x even mean?
  • Vibes — execs feel like we shipped 3x the value
  • Metrics — DORA numbers, PR throughput, features shipped
  • Outcomes — users get more value more quickly
  • All three matter. None is sufficient alone.
  • Risk: 10x some metrics while seeing modest gains on outcomes
// intro
AI compresses the 20%. The hard part is the 80%.
  • Writing code was always 20% of the work
  • Testing, refining, validating — always 80%
  • 10x on 20% doesn't get you to 3x overall
  • This talk is about the 80%: spec quality, review process, architectural discipline
// act-1: ai is an amplifier
Act 1
AI is an
Amplifier
// act-1: ai is an amplifier
AI is an amplifier.
  • DORA 2025 — 5,000 professionals, 100+ hours qualitative research
  • Magnifies the strengths of high-performing orgs
  • Magnifies the dysfunctions of struggling ones
  • 92% of devs use AI monthly — but outcomes are splitting
  • High-performing orgs: 50% fewer customer incidents. Struggling orgs: 2× more. Same tools.
  • "AI is moving organizations in different directions." — Laura Tacho, DX
// act-1: ai is an amplifier
What is AI about to amplify at your org?
  • Good architecture → great architecture
  • Bad architecture → messy, sprawling code generated at scale
  • Good specs → accurate, testable implementations
  • Vague specs → confidently wrong code that passes tests you didn't write
  • Consistent patterns → idiomatic output on the first try
  • Three divergent patterns → agent picks whichever it saw last, tech debt at speed
// act-1: ai is an amplifier
The velocity gains were temporary. The tech debt was permanent.
  • CMU — 807 repos after Cursor adoption
  • 281% increase in lines added, month one
  • Month two: velocity gains gone, back to baseline
  • Static analysis warnings +30%, code complexity +41% — permanently
  • Complexity increase → ~65% drop in future velocity
  • "Slop creep" — Boris Tane: individually reasonable, collectively destructive
// act-1: ai is an amplifier
Taste × Discipline × Leverage — multiplicative, not additive
  • Taste — when generation is free, knowing what's worth generating is the scarce skill
  • Discipline — specs before prompts, tests before shipping, reviews before merging
  • Teams skipping discipline are reporting production disasters — Cortex 2026: incidents/PR up 23.5%
  • Leverage — small teams, stacked PRs, agent orchestration, design engineers eliminating handoffs
// act-1: ai is an amplifier
This isn't theoretical.
  • Amazon retail: spike in outages from AI-assisted changes → senior sign-off mandate for junior/mid engineers
  • Anthropic's own website shipped a basic regression affecting every paying customer — 80%+ generated with Claude Code
  • "I don't think we're even trading this off to move faster. I think we're moving at a normal pace." — OpenCode CEO
  • Discipline is what separates teams that sustain from teams that crash
// act-1: ai is an amplifier
Best practices matter more with AI, not less.
// act-2: code quality is harness quality
Act 2
Code Quality is
Harness Quality
// act-2: code quality is harness quality
The bottleneck is the environment, not the model.
  • OpenAI harness engineering: 3 engineers (later 7), ~1M lines, 1,500 PRs, zero hand-written code
  • When the agent fails: "what capability is missing, and how do we make it legible and enforceable?"
  • Failures are harness problems, not prompt problems
// act-2: code quality is harness quality
Refactoring is AI infrastructure — not cleanup.
// act-2: code quality is harness quality
RIVET: backend layered architecture
  • Three divergent patterns → one: DB → Model → Repository → Service → Controller → Router
  • Each layer: one clear responsibility, one clear contract with adjacent layers
  • Defense-in-depth permissions at router, controller, and service levels
  • Agent can't accidentally create an unprotected endpoint — the pattern enforces security
  • Maps closely to OpenAI's enforced architecture. Not a coincidence.
// act-2: code quality is harness quality
RIVET: frontend state migration
  • Legacy: load full app state upfront, apply incremental updates
  • Problem: sluggish loads, global state — hard for humans and agents
  • New: fetch data when and where needed, local component state, explicit dependencies
  • Smaller blast radius per change — agent can work in one area without breaking adjacent ones
  • Two goals: better UX for users + a codebase Claude can work in safely
// act-2: code quality is harness quality
Mechanical enforcement > documentation
  • OpenAI: encoded architectural rules as custom linters, not prose
  • Linter error messages written as remediation instructions — every violation teaches the agent how to fix it
  • Rules that feel heavy in human-first workflows become multipliers with agents — apply everywhere at once
  • Documentation drifts. Lint rules don't.
// act-2: code quality is harness quality
Rules without refactoring are aspirations, not enforcement.
// act-3: context engineering
Act 3
Context
Engineering
// act-3: context engineering
From prompting to context engineering
  • Prompting — telling an AI what to do in a single interaction
  • Context engineering — building the persistent, structured context that makes an agent reliably useful over time, across a whole team
  • CLAUDE.md is where this starts — but the real harness extends into linters, tests, CI gates, observability, and automated maintenance
// act-3: context engineering
CLAUDE.md: map, not encyclopedia
  • A large instruction file actively degrades agent performance — crowds out the task
  • 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
// act-3: context engineering
Rules and hooks
  • Rules — constraints the agent follows automatically: coding standards, naming patterns, always/never
  • Hooks — automated actions triggered by agent behavior: linters after generation, auto-formatting
  • Together: continuous mechanical enforcement, everywhere at once, without human attention
  • At RIVET: rules in progress, hooks on the roadmap
// act-3: context engineering
PILRs: context that accumulates
  • Persistent Indexed Learning Repos — structured knowledge bases the agent can reference
  • index.json maps topics → documentation files. Agent loads what it needs, not everything.
  • PILRs grow as the agent works — organic accumulation vs. upfront documentation
  • Goal: "generic coding assistant" → "specialist who knows how our platform is built"
  • Addresses the cold-start problem. You don't have to document everything upfront.
// act-3: context engineering
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
github.com/obra/superpowers
// act-3: context engineering
PILRs at Signal Advisors
  • MCP server loads PRDs & tech specs as context for agents working across multiple code repos
  • Agents get the right domain knowledge without it living in any single repo
  • Same pattern, different delivery mechanism — the PILR principle scales beyond local files
// act-3: context engineering
Indexer — an AI-powered codebase index generator using Tree-sitter AST parsing, code skeleton extraction, and PageRank-based repo mapping
Indexer
github.com/AndyInternet/indexer QR code
github.com/AndyInternet/indexer
// act-3: context engineering
Skills: making agent usage programmable
  • Reusable packaged workflows — complex multi-step processes become a single command
  • At RIVET: /review-pr, /implement, /update-project-docs, bug-fix skills, PILR builders, commit rewriter
  • Skills are team assets: versioned, shared, improved over time
  • This is where the promises of GenAI for coding start to make real sense
// act-3: context engineering
The /implement skill
  • PRD + Tech Spec → Project Plan (reviewed before code is written)
  • Project Plan → Test Plan: automated tests + manual testing steps
  • Agent writes code, executes test plan via cURL + Claude Chrome Extension
  • Screenshots → gitignored folder → referenced in pr.md
  • Honest: ~25% success rate on screenshots. Worth it anyway.
  • Start with /review-pr or /commit — not /implement
// act-3: context engineering
The support bot agent: "re-purchased half a developer"
  • Before: ticket → ~20 min understanding + ~20 min finding the code + fix time
  • Now: agent triages, reproduces, proposes fix or surfaces what the engineer needs
  • Uses a PILR — persistent directory of previously solved problems + index
  • Result: support backlog per two-week shift dropped significantly
  • Doesn't solve novel architectural bugs — good at "we've seen this pattern before"
  • Why it works: the harness came first. Without CLAUDE.md, the agent just guesses.
// act-3: context engineering
Spec-Driven Development
  • Specify → Plan → Tasks → Implement
  • Expanded the "safe delegation window" from 10–20 minute tasks to multi-hour feature delivery
  • The harness makes the agent reliable. SDD makes the work product predictable.
  • Forces hard thinking to happen before implementation, not during it
  • Claude Code's plan mode is built for this workflow
// act-3: context engineering
4-layer review
  • 1. Architectural review — human, before code is written
  • 2. Automated review — /review-pr skill, SonarCloud, Copilot review, 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)
  • Unsolved: without ephemeral envs, reviewers contend for shared test servers — gets worse as PR volume increases
// act-4: multi-agent reality check
Act 4
Multi-Agent
Reality Check
// act-4: multi-agent reality check
The promise: a fleet of AI collaborators
  • One agent orchestrates many — background agents on separate branches in parallel
  • An engineer manages a fleet, not a single assistant
  • Cursor is doing this. Ramp. Stripe. Large AI-native companies.
  • Orchestration platforms emerging: Coder and others — provisioning environments, coordinating agents across a codebase
// act-4: multi-agent reality check
Where we actually are
  • ✓ GitHub Copilot cloud agents
  • ✓ Support bot agent — and it's working
  • ~ Running 1–3 local tasks at once — sometimes
  • "I've read sensational claims from Ramp and Stripe. We don't have things tuned to that level."
  • "I haven't figured out how to multiplex my brain yet."
// act-4: multi-agent reality check
What actually happened when I tried
  • Context debt: you need rich structured context in place — we weren't there yet
  • Review bottleneck: more agents = more PRs = worse pile-up
  • Cognitive load: managing multiple agent threads is a different skill, not just more of the same
  • Not that the tools don't work — we hadn't earned the right to use them yet
// act-4: multi-agent reality check
We're in the j-curve. The dip before the improvement.
  • The things that are working are working because of harness investment
  • Multi-agent nirvana requires the full harness first — we're not done building it
  • This is a living experiment, not a solved problem
  • Anyone who tells you they've fully solved this is either Cursor or not telling the whole story
// close
Monday morning
  • Write a CLAUDE.md for one repo: tech stack, architecture in two paragraphs, three things you wish every new engineer knew
  • Try one skill — start simple: /review-pr or /commit, not /implement
  • Identify one refactor that would make your codebase more agent-navigable — put it on the roadmap
  • None of this requires a BHAG or an executive mandate
// close
Further reading
  • DORA AI Capabilities Model (2025) — google.com
  • Harness Engineering — OpenAI (Feb 2026)
  • Building an Elite AI Engineering Culture — CJ Roth (Feb 2026)
  • Are AI Agents Actually Slowing Us Down? — Pragmatic Engineer (Mar 2026)
  • Speed at the Cost of Quality — Carnegie Mellon (2026)
  • Full blog post with all references: detroitdevelopers.com
  • Superpowers plugin for Claude Code: github.com/obra/superpowers
Blog post QR code
detroitdevelopers.com/blog/agentic-coding-advanced-guide/
// close
Questions
detroitdevelopers.com QR code
detroitdevelopers.com