Matthew Dresden
DevBench
Matthew Dresden · AI-Native Software Delivery

From velocity
to trust.

Autonomous software delivery is here. The hard part isn't speed anymore — it's accountability. This is how DevBench builds enterprise trust into every step, and how Matthew Dresden helps you adopt it.

Built on Claude 242 tasks in one public run 99.93% cache-hit efficiency Public proof: caylent-solutions/kanon Many more delivered by Matthew Dresden
A walkthrough for engineering leadership — press → or Space to begin.
The shift

Last year we chased velocity.
That problem is largely solved.

When executing a backlog becomes trivial, two questions decide who wins: what should we build, and can we trust the system that builds it?

The old way
A team of 10

Coordination overhead, handoffs, and inconsistent quality — speed bought with headcount.

The new shape
1 architect + a trusted digital team

A person defines the work. A disciplined team of single-purpose agents executes it — with gates a human team would recognize.

Before any agent runs

Garbage in, garbage out — engineered out.

The hardest part of autonomous delivery isn't the agents. It's giving them work so unambiguous it can't be misread. DevBench makes that the easy part — with guided skills.

One unit of work, end to end

How a single task gets done.

Click any node to see what it does — and any 🛡 guard to see how trust is enforced at that step.
step 🛡 deterministic guard complete
Why you can trust it

Trust isn't a promise. It's the architecture.

Six design choices, working together — the answer to "why would I let AI touch production?"

Real-time visibility

You see everything. You babysit nothing.

A real, completed DevBench run — the public kanon project.

0%Complete
242 / 246 tasks · 468 / 479 work units
1 declined · 0 blocked at finish
0%
Cache-hit rate
~$0
Total cost, full project
0 h
Wall-clock span
0 min
Avg / task (10.6 recent)
0 M
Tokens (617M cached)
0
Stories/features/epics auto-rolled
Backlog — live replaycaylent-solutions/kanon · feat/kanon-deps-work-2026-05
Real-time alerts — it tells you, you don't check on it
Work unit done Needs an operator PR opened PR merged on green CI !CI failed — auto-retrying
Smart block triage — most blocks recover on their own

DevBench classifies every block. Most recover automatically; only genuine operator-action blocks notify a human — and it keeps working other tasks while you resolve those later. One block never stalls the run. And it's fully configurable: be as hands-off or as hands-on as you want.

reviewer-feedback retry in-scope amendment CI-failure retry PR-bot feedback retry blocker-resolver → new tasks
Honest by design: every number is from one real, public run — verifiable commit-by-commit on GitHub. Nothing here is simulated.
Breadth

It's backlog-driven — so it isn't only for code.

Whatever you can express as a backlog with clear acceptance criteria, it can execute and prove.

Where Matthew Dresden comes in

We don't just have the tools.
We help you build the muscle.

Adopting AI-native delivery is a change in workflow and trust, not just tooling — codebase-aware definitions of done, new review practices, and a "trust, but verify" posture. That's the engagement.

1

AI-SDLC Workshop

Bring a real idea. Over a couple of days we run it through the process and train your team on it.

2

Guided engagement

We stand the practice up inside your org — with the gates, audit, and judgment model that make it safe.

3

Your team owns it

You keep the architects and the judgment. The execution layer scales to whatever the backlog demands.

Matthew Dresden — Bringing Out the Best in People and Technology.
Presenter note
1 / 8
The differentiator

Many partners can say "AI-SDLC." Fewer can say why theirs is trustworthy.

Each choice below is grounded in public evidence — not marketing.

Roadmap — planned, not yet shipped

Where DevBench is heading.

DevBench evolves to meet customer needs — especially for regulated environments. These are in design today, shown so the conversation can include where we're going. They are not current capabilities.

Shipped capabilities are on The Walkthrough. This page is intentionally forward-looking and labeled as such.