← All capabilities
Maintenance
Phase: ImproveKeep what works today working when the world shifts underneath it.
Maintenance keeps production agents alive when what’s underneath changes: models that update or retire, systems that change their API, business processes that evolve. AI isn’t a project you ship and forget; it’s a system you operate.
What it includes
- Watch provider model changes (versions, deprecations) and their effect.
- Update integrations when connected systems change.
- Detect drift: when quality drops because the world changed, not the code.
- Manage versions and be able to roll back if a change makes things worse.
When you need it
- An agent that was working started failing without anyone touching anything.
- The provider is retiring the model you use and there’s no plan.
- Every change in a connected system silently breaks the integration.
How it’s measured
- Uptime of production agents.
- Incidents from drift or external change vs. total.
- Time from an external change appearing to it being absorbed.
Common mistakes
- Treating deployment as the end of the project and assigning no maintenance owner.
- Not watching provider deprecations until the agent stops working.
- Switching models without re-running the evals and trusting it "will behave the same".
FAQ
- Why does a working agent start failing on its own?
- Because the world shifts underneath: the provider updates the model, a connected system changes its API, or the input data evolves. That’s called drift, and catching it early is part of maintenance.
- How much maintenance does production AI need?
- More than people expect. A production agent is living software with external dependencies that change. Without maintenance, reliability degrades on its own over time.
Related capabilities
Missing any of these?
Start with an assessment: which capabilities you have, which you’re missing, and where it pays off most to start.
Talk to us →