Blue-green deployment is a release method that keeps two identical production environments, blue (live) and green (new). The new version is deployed to green, checked, and then all traffic is switched to it at once, with blue kept ready so you can switch back immediately. For AI agents, the switch moves both the version and its authority.
Why does blue-green deployment matter for agents?
Agent changes are not only code changes. A new brief, a new model, a new source or a new tool connection can each change behaviour. Blue-green gives you a clean cut-over and a fast route back: if the new version misbehaves, traffic returns to the old one in one step, without a rebuild.
The agent-specific catch is authority. The green version is a different agent in substance, even if it carries the same name. It needs its own certification at the depths it will act, before the switch. Otherwise you have a fast switch to an unproven agent.
How does it work in practice?
- Prepare green. Deploy the new version with its brief, sources, tools and model.
- Certify green. Replay it against the customer's own cases, up to the depth it will hold. Green does not inherit blue's certificate.
- Check connections. Confirm green reaches the same systems with the same scoped identity and limits.
- Agree the switch. Name who approves it and the conditions for switching back.
- Switch. Route all new tasks to green. Let blue finish in-flight tasks, or hand them over cleanly.
- Watch. Hold rates, refusals, reviewer edits and errors for a set period.
- Retire or return. If green holds, retire blue. If not, switch back and investigate.
The AI agent deployment plan template has fields for each step, including the switch-back criteria.
What does an example look like?
A UK retail bank's operations team runs an agent that drafts responses to direct debit queries. A new version adds a source for scheme rule updates. The team deploys it as green, certifies it on several hundred past queries, and switches over at the start of a quiet morning. Two hours in, reviewers report that green cites the new source for cases it does not cover. The team switches back to blue in a single step, and green goes back for a narrower source scope. The kind of operations work involved is described in AI agents in UK retail banking operations.
How is blue-green different from a canary release?
Blue-green switches all traffic at once and relies on the fast switch-back for safety. A canary release sends a small share of traffic to the new version first and widens it gradually. Canary limits how many cases a bad version can touch; blue-green limits how long it can touch them. For agents that act at depth 3 or above, a canary is often the safer first step, with blue-green used for the final cut-over.
In-flight work is the other difference to plan for. Agent tasks can span hours or days, such as a case waiting on a customer document. Decide whether those finish on blue or move to green, and record which version handled each step.
Which terms sit close to it?
Rollback: returning to the previous version, which blue-green makes quick. Feature flag: switching a capability on or off without a release. Change management: the approval around the switch. A board weighing a significant agent change may ask for the switch-back plan, which a board paper on AI agent deployment can include. See our guide to deploying AI agents for the wider picture.
Questions people ask
Does blue-green double the cost of running an agent?
Only during the change window, and mostly in idle infrastructure. The model calls follow the traffic, which goes to one side at a time.
Can blue and green share a certificate?
No. A certificate covers the agent as certified. A changed brief, model or source means a different agent, which needs its own evidence.