Ops Conductor

• 

  1 messy brief → 14 issues with labels, dates, dependencies, and batched execution plan

• 

Inspiration

I run six ventures, take gig shifts, co-parent a teenager, and teach AI workshops. My brain holds more open loops than any to-do list can manage. I built an AI life operating system called Douglass to handle it — and realized the same pattern that organizes my life (decompose, classify by activity type, sequence to minimize context-switching) could help any developer manage complex projects inside GitLab.

What it does

Ops Conductor is a GitLab Custom Agent that acts as a DevOps chief of staff. Give it a project brief — "launch user onboarding by Friday" — and it:

Decomposes the brief into a full Work Breakdown Structure Creates GitLab issues for each task with descriptions, acceptance criteria, and time estimates Classifies every task by cognitive activity type: TECHNICAL, COMMUNICATION, CREATIVE, LOGISTICS, ADMIN Detects dependencies between tasks Posts a sequenced execution plan grouped by activity type — so all coding happens in one block, all communication in another, all admin at the end

The key insight: cognitive-mode batching. Developers lose 40-60% of their productivity to context-switching. Alternating between writing code, sending emails, and doing admin wastes the day. Ops Conductor eliminates this by sequencing tasks so similar types of thinking happen together.

How we built it

GitLab Duo Agent Platform (Custom Agent) Anthropic Claude via GitLab's built-in model access Built-in tools: Create Issue, List Issues, Create Work Item System prompt encodes WBS decomposition logic, 5 activity types, and dependency detection No external API calls — 100% platform-native

Challenges we ran into

The biggest challenge was working within the GitLab Duo Agent Platform's constraints. My original plan was a standalone Python agent calling Claude's API directly, but research showed that approach would score poorly on platform leverage — the #1 judging criterion. Pivoting to a Custom Agent meant the entire agent had to live in a system prompt and built-in tools, with no external code. That forced me to encode 22 years of operations decomposition experience into prompt engineering rather than application logic. The second challenge was cognitive-mode batching itself — teaching Claude to classify tasks not by subject matter but by the type of thinking they require. "Write API docs" is TECHNICAL by subject but CREATIVE by cognitive mode. Getting that distinction right in the prompt took multiple iterations.

Accomplishments that we're proud of

The cognitive-mode batching approach is genuinely novel — no other project planning tool sequences tasks by the type of thinking required. This comes from real experience managing 140+ person teams across NFL venues and music festivals, where the difference between a smooth operation and chaos is whether people are switching between logistics, communication, and execution, or batching similar work together. I'm also proud that this agent is one module of Douglass (getdouglass.app), an AI life operating system I built and use daily to manage six ventures, gig shifts, and co-parenting. The agent works because the decomposition logic is battle-tested on my own life.

What we learned

The biggest insight was that project decomposition is a reasoning problem, not a coding problem. The entire agent is a system prompt and tool configuration — no Python, no JavaScript, no external services. Claude's reasoning capability does the hard work of understanding a project brief, identifying the tasks, detecting dependencies, and sequencing by cognitive mode. The platform's built-in tools handle the execution.

What's next for Ops Conductor

Ops Conductor is one module of Douglass (getdouglass.app), an ambient AI life operating system I use daily. The cognitive-mode batching approach will expand to include:

Automated daily standups that scan all open issues and generate prioritized briefings CI/CD pipeline monitoring that flags failures and creates remediation issues Sprint planning that decomposes epics into batched work blocks

Built With

• ai-agents
• anthropic-claude
• devops
• gitlab-duo-agent-platform