Introducing Kairon Runtime 3.0

What is in Runtime 3.0

Kairon Runtime 3.0 is available today for every workspace. It is the largest change to how agents run on Kairon since we launched, and almost all of it came from one request we heard over and over: make agents behave the same way in production as they did in testing.

There are three headline changes: one API for every tool connector, durable steps that survive restarts, and a rebuilt trace viewer. Existing agents keep working without changes. You can opt in per agent, at your own pace.

400+ connectors, one API

Until now, each connector had its own small quirks: different error shapes, different pagination, different ways to report a rate limit. Agents had to learn those quirks, and so did the people writing their prompts.

In 3.0, every connector, from Payly and Bookwise to Relay and Cloudly, speaks the same interface.

• One error format with a clear retryable flag, so agents know when to wait and when to re-plan.
• One pagination model, so a list of invoices and a list of bookings behave the same way.
• Idempotency keys by default on every write, so a retried step never charges, books or sends twice.

In beta, teams that moved to the unified API cut connector-related prompt instructions by an average of 60%.

Durable steps

Long tasks used to be fragile. If a worker restarted halfway through a twelve-step run, the run failed and someone had to start it again. That was rare per run and common across thousands of runs.

Every step in 3.0 is checkpointed before and after it executes. If anything interrupts a run, it resumes from the last completed step, with its context and approvals intact. A run waiting three days for a human approval now costs nothing while it waits and picks up exactly where it stopped.

Our goal for 3.0 was boring reliability. An agent should never have to do a step twice because our infrastructure blinked.

During the beta, durable steps recovered 99.97% of interrupted runs without any human help.

A trace viewer rebuilt from scratch

The new viewer organises each run around decisions rather than raw events. For every step you see what the agent knew, what it chose and what happened. You can compare two runs side by side, and approvals appear in the timeline where they happened. We wrote about the thinking behind it in traces you can actually read.

How to upgrade

1. Open an agent’s settings and switch the runtime to 3.0.
2. Run its eval suite. The runtime change is recorded as a new version, so you can compare results directly.
3. Promote it through shadow mode and canary as usual.

Runtime 2.x remains supported until the end of March, with security fixes after that through June. Pricing does not change. Full details are on the changelog, and if you have questions, talk to our team.