The demos were not the work
Most AI coding demos still feel detached from real software delivery. You ask for code, receive a snippet, paste it somewhere, and hope it fits.
That can be useful, but it skips the structure where the work actually happens: issues, labels, branches, commits, merge requests, checks, and review.
Connect the agent to the workflow
I connected OpenCode to GitLab through a custom Model Context Protocol integration. The important part is not the protocol itself. It is that the agent can work with the same delivery objects my team already uses.
I can ask it to turn a request into a ticket, make a focused change, run checks, push a branch, and prepare a merge request with the context needed for review.
Let the machine handle the mechanics
The useful automation is not a dramatic one-shot feature build. It is the chain of small jobs around the code: finding the right files, keeping requirements visible, running the boring checks, and recording what changed.
That lets me step away instead of driving every tool call. When I return, the work should be waiting in a familiar place with a diff I can inspect.
Keep the decision human
I still review the merge request, check whether the implementation makes sense, and verify that it solves the actual problem. The agent can move work forward, but responsibility does not move with it.
For me, that is the useful shift. AI stops acting like autocomplete and starts operating like an engineer inside a controlled workflow, while a human remains responsible for direction and the final decision.
Practical beats magical
I am not interested in pretending the system is autonomous or replacing the review process. I want fewer repetitive steps, less context switching, and a clean trail of what happened.
That is less exciting than the usual AI promise. It is also much closer to something I can trust in daily software work.