Agent Workflow Review
Two weeks. Fixed fee. You keep the fixes.
A close look at how your team actually uses Claude Code, Cursor, Copilot or Codex: where the agents help, where they break, and why. It ends with the top fixes already made.
Who it's for
Engineering teams of roughly 5 to 50 developers that already pay for coding agents and expected more from them. The person who calls is usually a CTO, VP of Engineering, head of platform or developer experience, or a founder who still reviews code.
What I look at
- Read access to one or two repositories
- Your agent setup: CLAUDE.md / AGENTS.md, skills, MCP servers, hooks, CI wiring and permissions
- A sample of real agent sessions, chosen by you
- 60–90 days of pull-request and review history
- What you spend on AI tools
- 45-minute conversations with the engineering lead and three to five engineers
What you get
- A findings memo, ten pages at most, ranked by impact: where agents help, where they break (context loss, instruction bloat, review bottlenecks, unsafe tool permissions, cost leaks) and why.
- A pull request with the top three to five fixes already implemented, for your team to review and merge.
- A baseline scorecard to measure future changes against: spend per developer, review time, PR size, rework signals.
- A 60-minute session walking your team through all of it.
- A scoped proposal for what comes next, if anything should.
How the two weeks run
- BEFORE DAY ONENDA signed, access set up, the session sample exported.
- WEEK ONEReading the setup, the sessions and the history; the interviews.
- WEEK TWOFindings, the fix PR and the scorecard; the team walkthrough on the last day.
Not included
- Production deploys
- More than two repositories
- Vendor selection beyond a recommendation
- More than one on-site session (Denver area)
- Support after delivery (that's the fractional engagement)
Access and data handling
- NDA first. A mutual NDA, yours or mine, signed before any access.
- Least access. Access that can read the agreed repositories and open a pull request, and nothing that can merge or deploy. It's revoked the day we finish. No production systems, customer data or secrets; I don't need them.
- You choose what I see. You export the session sample and strip anything sensitive first. If a secret turns up anyway, I tell you and delete it.
- Fixes arrive as a pull request. Your team reviews and merges. I never merge, deploy or touch production.
- No training on your code. Your material is only processed by AI services whose terms exclude training on it. If you'd rather nothing leave the building, I work only with local models on hardware you own, and setting that up can be part of the engagement.
- Deleted at the end. Within 30 days of delivery I delete your material and confirm in writing. The deliverables are yours.
- Nothing published without you. The engagement is confidential. Any case study is written with you and published only with your written approval.
Price
A fixed fee, agreed before we start and quoted on the check-up call. Half on signing, half on delivery.
FAQ
- Which tools do you work with?
- Whatever your team uses: Claude Code, Cursor, Copilot, Codex, or open-weight models run locally. The failure modes (context, instructions, review, permissions, cost) are mostly the same across tools, and I'll say up front if something is outside what I know.
- Remote or on-site?
- Remote by default. One on-site session in the Denver area is included.
- Do you need production or customer data?
- No. See access and data handling.
- What happens after?
- Most teams take the fixes and run. If the memo points at more work, the proposal scopes it: a three-to-six-week build-out, a fractional lead about a day a week, or a specific build. No obligation.
- How big a team is this for?
- Roughly 5 to 50 developers. A larger organization can start with one team.
Are your coding agents paying for themselves yet?
Book a free 30-minute check-upphil@basswoodcreative.com