Most teams I talk to are in the same place. Everyone has an AI coding tool, a few people get a lot out of it, most get a little, and nobody can say whether the pull requests got better or just bigger.
The gap is rarely the model. It is everything around it.
#What I have learned doing this for real
I build LLM Gateway with coding agents every day, and the setup that makes their output trustworthy lives in the repo next to the code:
- Repo-local skills that encode the conventions an agent would otherwise guess at: how to add a provider, write a migration, open a pull request, cut a changelog entry. They are reviewed like code.
- A skill that audits the other skills, because agent instructions fail silently by going stale, not loudly by breaking.
- Scheduled routines that run recurring work unattended and open pull requests with fixes, instead of reports nobody reads.
- An isolated stack per worktree, with its own ports, Postgres and Redis, so several agents work in parallel without colliding.
- Deliberate context hygiene. Deny-lists keep build output and generated files out of the context window. This does more for quality than switching models.
At Axis Communications, where I consult through Sigma Technology Group, I have shared the same workflows with the team: GitHub Copilot, reusable skills, MCP integrations and agent orchestration.
I also build the tools on the other side of this. DevPass gives a team one key and a flat price across models in Claude Code, OpenCode, Cursor and Copilot, and DevPass Code is our own open-source terminal coding agent.
#What you typically hire me for
- Finding out, honestly, whether agents are helping your team today and where they would help most.
- Setting up skills, MCP servers, sandboxes and review gates in your repositories.
- Training engineers on your own codebase, not on a toy app.
- Choosing tools and model access without getting locked into one vendor.
- Writing the policy for what agents may and may not touch, in language engineers will follow.
If your team has the licences but not the results, tell me where you are and I will tell you what I would do first.