Agent engineering for teams on Claude Code, Codex and Cursor

Your engineers already pay for coding agents. I turn that into a system.

I'm Derek Keith, a software engineer with 20 years of shipping production software. Today I build a full trading platform mostly through AI agents, with shared skills, hooks, MCP servers and review gates doing the heavy lifting. I set up the same system on your repo, and we measure whether it works.

boot.log

tty0 · profitabul
  • [ OK ]skills: 75 shared and versioned
  • [ OK ]mcp: ~50 tools connected
  • [ OK ]memory: ~250 entries synced
  • [ OK ]tests: 3,953 files guarding every change
  • [ OK ]ci: 10 workflows, review gates armed
  • [ OK ]agents: claude, codex
  • [ WAIT ]your repo: audit pending
  • $

■ The gap

Seats are not a system

Most teams bought licenses and left the rest to each engineer. The result is uneven use, no shared rules, and gains that are hard to see.

24%

of employers have published approved AI tools or formal guidance for engineers

Stack Overflow pulse survey, May 2026

21%

of AI coding licenses go unused or underused

Opsera 2026 benchmark

~8%

median gain in pull-request throughput; teams that invest in rollout and training do far better

DX

■ Insert coin

Hand me a task. Watch it ship live.

A fixed-scope task, done on the dev stream with my agents and reviewed by me before it reaches you. You get more than AI output: you get an engineer's judgment on every line.

Now playing● Queue open

Your task could be next on stream.

Shipped on streamPR · time
  1. 1ST --------------:--
  2. 2ND --------------:--
  3. 3RD --------------:--

The board starts with the first delivered task.

You get

  • ✓A pull request with tests
  • ✓My written review notes
  • ✓The agent setup used to build it
  • ✓The recording

House rules

  • Fixed scope: we agree the task and acceptance criteria before anything is charged.
  • I accept the task first. You only pay once it's accepted.
  • If the time box ends first, you get the partial PR and a written plan for the rest.
  • Public and sandboxed repos run on air. Private code runs off air: the stream shows the board, CI and timer, never the code.

■ Offers

Fixed fees, published up front

For teams: every engagement starts with an audit, so you know what you're buying before you commit to more.

Agent readiness audit

Start here

1–2 weeks, fixed fee

$7,500 – $15,000

Where your team stands today, and a 90-day plan to get real value from the seats you pay for.

  • Review of current agent use, spend and configuration
  • CLAUDE.md / AGENTS.md, permissions and governance assessment
  • Baseline metrics: cycle time, PR throughput at constant review quality, agent cost per engineer
  • Written 90-day plan you can run with or without me
  • Startup version from $2,500, credited toward follow-on work

Team workshop on your repo

1 day, or 3 days with follow-up

From $6,000

Hands-on enablement on your own codebase, not toy examples.

  • Up to 12 engineers, remote or on site
  • Working skills, hooks and review loops committed by the end
  • 30 days of follow-up support with the 3-day version

Agent platform sprint

3–6 weeks, fixed scope

$25,000 – $60,000

The shared system your whole team works through.

  • Shared skills and hooks for your stack
  • MCP servers connecting agents to your internal systems
  • Memory conventions, CI review gates and a written agent policy
  • Before-and-after measurement and a clean handover

Fractional AI-engineering lead

Monthly retainer

From $6,000 / month

Ongoing reviews, new tools and upkeep as models and vendors change.

  • Advisory or hands-on tiers
  • New skills and MCP tools as needs come up
  • Model and vendor change management

■ Proof

I run this system every day

Everything I install for clients comes from how I build my own products.

A production trading platform, built by one engineer working through agents

Since September 2025 I've built Profitabul almost entirely through Claude Code and Codex. The point isn't speed for its own sake. It's that the agent setup keeps the work disciplined: every change goes through tests, review gates and CI before it ships.

merged pull requests
1,175
✓ merged pull requests
test files
3,953
✓ test files
database migrations
382
✓ database migrations
shared agent skills
75
✓ shared agent skills
tools in its MCP server
~50
✓ tools in its MCP server
surfaces: web, iOS, Android, desktop, browser extension, CLI
6
✓ surfaces: web, iOS, Android, desktop, browser extension, CLI

bleh.xyz

Open-source market-data infrastructure in Rust, from first commit to live in five days

An open market-data mesh with clean lineage. Agents wrote the code under the same kind of rules and CI gates I install for clients. Every commit is public, so you can inspect how the work was done.

first commit to live site
5 days
✓ first commit to live site
Rust crates
12
✓ Rust crates
commits co-authored with Claude
180 of 182
✓ commits co-authored with Claude

A self-directing live stream

built on stream

Built by parallel agents in one morning

A broadcast that shows agent work as it happens and switches scenes on its own when a pull request merges or a deploy lands. It filters secrets and sensitive details before anything reaches the screen, and fails safe when its data feed drops.

commits between 6:35 and 11:50 one morning
39
✓ commits between 6:35 and 11:50 one morning
manual scene switches
0
✓ manual scene switches

■ Run of show

Measured on your repo, owned by your team

  1. 01Start

    Audit

    I review how your team uses agents today and take baseline measurements on your repo.

  2. 02Next

    Install

    Skills, hooks, MCP servers, review gates and a written policy, fitted to your stack and your rules.

  3. 03

    Measure

    Before-and-after numbers on cycle time, throughput at constant review quality, and agent spend. No multipliers, just your data.

  4. 04

    Hand over

    Your team owns everything I build. It lives in your repo, documented, with no lock-in.

■ About

20 years of shipping, now through agents

Derek Keith
Live with DEREK● 20 yrs

I've led and built software for healthcare, government technology and marketing automation companies: identity and authentication for hospital services, React Native patient apps, procurement platforms and greenfield Next.js products.

Regulated environments taught me to care about what an auditor can read. That's the standard I bring to agent workflows: clear rules, review gates, and a record of what changed and why.

$ ./start

Find out where your team stands

Start with a fixed-fee audit. You get a baseline, a plan and a clear picture of what your agent seats are worth, whether or not we work together after.