DayOne Agent

What inspired me

When I started working a few years back I joined a new team and nobody had time to walk me through the codebase. There was a wiki that hadn't been updated in months. I spent two weeks clicking through files trying to figure out what actually mattered.

The repo knows everything about itself. The commit history, the hot files, the people who touch things most. Why does nobody use that to help new people in?

When I found the GitLab Duo Agent Platform I immediately thought: this is the thing that could fix it.

How we built it

We started with the flow trigger — member_added felt like the right moment. Not when someone requests access, not when they push their first commit. The moment they join is when the anxiety starts.

We built a custom agent in the AI Catalog, gave it access to the repository tools, and wrote a system prompt that told Claude exactly what a good first-day guide looks like. The AGENTS.md file does a lot of the heavy lifting here — it defines the tone and output format so the guide doesn't read like a generated document.

The hardest part was actually the output. Getting Claude to write something that sounds like a senior developer wrote it for a colleague, not a bot filling a template — that took more iteration than anything else in the build.

What we learned

We underestimated how much the platform context changes what's possible. Having an agent that already understands the relationships between things in a repo is a completely different starting point than reading files you point it at.

We also learned that AGENTS.md is more powerful than expected. It's not just context — it's how you give an agent a personality and a set of standards it actually follows consistently.

Debugging flow inputs was a key lesson — the context:project_id input must be explicitly declared for the workflow to execute correctly. Missing this caused sessions to close immediately without running.

Challenges

Two things slowed us down.

First — flow session state. Every trigger was resuming an existing workflow instead of starting fresh. The root cause turned out to be a missing project_id input in the flow definition. Once declared, the flow ran correctly every time.

Second — tone. Every time we tested the agent the output was technically correct but felt robotic. We rewrote the system prompt multiple times before it started producing guides that felt genuinely welcoming rather than procedurally generated.

What's next

A 7-day follow-up agent that checks in with the new member — are the files we pointed you to still the right ones, did you find the people we mentioned, what's still confusing?

That second touchpoint is where the real value compounds.