Alex Vydrin

Agent-first development

Alex Vydrin

I set up agent-first development for software teams: from a written plan to production, with the review, CI and merge gates that make it safe.

Book a 30-min call

Senior backend engineer. Based in Barcelona, working remotely, happy to travel.

Who I am

I was one of the engineers who built and ran CRED's agent-first engineering system: plans that fan out into parallel agents across repos, AI review, self-healing CI, a review loop that fixes comments on its own, and auto-merge straight to production. I built the CI, review-loop, auto-merge and shared-skills layers myself, and worked on the plan-and-fan-out side every day. Now I set up the same pipeline for other teams, adapted to their stack and their pace.

The AI factory

Work moves from a written plan to production through coding agents. People decide what to build and where the gates sit. Agents do the typing, and every step leaves evidence behind.

  1. Plan

    A written plan with scope, repos and acceptance checks.

  2. Fan-out

    The plan splits into tasks that parallel agents run across repos.

  3. AI review gate

    Every pull request is reviewed automatically before a person looks at it.

  4. Review loop gate

    An agent fixes the comments and asks for review again.

    until no Critical or Major is left

  5. CI built for agents gate

    Fast, predictable checks, and infra failures rerun on their own.

  6. Auto-merge and deploy gate

    Merge rules decide what ships on its own and what waits for a person.

  7. Post-merge verification

    After deploy, the change is checked where it actually runs.

  8. Shared skills and handoffs

    People talk in Slack, agents leave evidence in Linear. Shared skills keep every agent working the same way.

gate a step that has to pass before work moves on.

Proof

CRED

Remote US enterprise AI and data platform

  • Review loop adopted by 7 engineers, about 1,700 runs.
  • Auto-merge rolled out to 5 repos, 3 of them deploy to production on merge.
  • CI for agents: compile from 28 min to 93 s, unit tests from a 45-min timeout to 154 s.
  • Bounded auto-rerun of infra failures on 1,400+ PRs.
  • Bot PR flood cut from 45 to 1 a day.
  • 88% of my PRs written with Claude Code by October.

Atmalogía

Education company, about 50 people. I am their part-time CTO.

  • A contractor quoted 115 to 200 hours for a site. I built the full scope in about one day with AI-assisted development, and it became the base for all company sites.
  • Built the company platform.
  • Built a payment-to-CRM gateway, now in production: 17 of 17 payments and zero duplicates on day one.

How we work

Audit

1 to 2 weeks

Interviews with your developers, a review of your repos, CI and how the team uses AI today. You get a concrete rollout plan with metrics.

Let's talk

Build

4 to 8 weeks

I set up the pipeline end to end: plans, agents, review loop, CI gates, merge rules and shared skills. Then I train the team to run it.

Let's talk

Retainer

A few hours a week

I stay on as your fractional AI-enablement lead, tuning the pipeline as your team and stack change.

Let's talk

Before engineering I spent six years as a journalist in Kyiv. It taught me to ask the right questions and explain complex systems simply.