Introducing Agent Versioning: ship faster, rollback faster

Every team that runs AI agents in production eventually hits the same moment: you ship a prompt change on Tuesday, quality drops on Wednesday, and you spend Thursday trying to work out which of the four edits you made was the culprit. Rollback means digging through commit history, re-pasting prompts, and re-deploying, assuming you committed the prompt in the first place.

The problem isn't that teams are careless. It's that agents didn't have the same deployment primitives software has had for decades. No diff. No version history. No one-click rollback. We built Agent Versioning to fix all three.

The second-order effect is on iteration speed. When rollback is painful, teams ship fewer changes. When rollback takes 12 seconds, teams ship boldly and recover quickly. Version safety doesn't just reduce incidents, it accelerates the entire improvement loop.

Every time you save an agent configuration, meaning the system prompt, retrieval settings, escalation rules or model parameters, a new version is created automatically. Versions are SHA-addressed, so identical configurations share the same hash. You can tag any version with a human-readable label and set any version as the active production version independently of the one you're editing.

Each version stores the full configuration snapshot: the complete system prompt, the model ID and temperature, the knowledge base revision hash, and the escalation rule set. This means a rollback restores everything, not just the prompt text, but the model and retrieval configuration that went with it.

Side-by-side eval runs are built into the version UI. Select any two versions, point them at a conversation set, and get a comparison table showing response quality scores, escalation rates, and latency. You can promote a challenger to production from the same view.

The diff view shows changes between any two versions as a structured diff: additions in green, removals in red, unchanged content collapsed. For long system prompts, the diff focuses on changed sections so you're not scrolling through 800 unchanged words to find the three that changed.

Rollback is a single button in the version timeline. Click the version you want, confirm, and within 12 seconds all new sessions are routed to that configuration. In-flight sessions finish on their current version, because we don't cut conversations mid-stream.

The next release adds A/B traffic splitting: route 10% of sessions to a challenger version, measure statistical significance on your quality metrics, and promote or discard automatically when significance is reached. No manual monitoring required.

After that, we're building version-aware alerting. If quality or escalation rate shifts by more than a configurable threshold within 30 minutes of a version deploy, you get an automatic incident notification with a link to the rollback button. We're targeting a workflow where a bad deploy is caught and rolled back before most customers experience it.