Kiwi agents: the new agent lanes, the bounded workflow | and where it still hurts

Hey,

Since the last thread we've rebuilt how agents actually run in Kiwi, not just what they can do. Here's what shipped, and the specific questions we can't answer from our own logs.

Two lanes, not one pool. The 14 agent templates (sales, marketing, PM, finance, HR, support, legal, ops…) run on a business model. Kai, the developer agent, now runs on its own code model with a bigger context window, its own pod, its own lifecycle. Different work, different engine, a sales follow-up and a refactor were never the same job.

Kai got a workspace. Local folder access from the browser, a fullscreen IDE (file tree, tabs, editor, save), and a floating Workspace window you can keep on a second screen with your todo list and the chat side by side.

The turn is bounded now. Multi-step requests get decomposed once into a plan instead of the agent improvising. Client-facing drafts (emails, files) get one critique-and-rewrite pass before they reach you. Every turn is traced end to end, so an action that fired can be tied back to the message that caused it. And there's a hard cap on steps per message : no runaway loops, by construction.

Actions are risk-tiered. Low risk runs. Medium runs and tells you. High risk waits for your approval, and we moved outbound webhooks and unatttended email sends up into that tier, because those are the ones that leave the building.

Everything runs on our own hardware. No Anthropic, no OpenAI, no third-party agent platform in the path anymore. Your CRM data doesn't get shipped to a US API to be summarized. The honest trade-off: the pod sleeps when idle, so the first message after a quiet spell waits for it to wake.

That trade-off is the first thing we want your read on:

- Cold start. Is waiting for a pod to wake on your first message acceptable, or does it kill the habit? Keeping one always warm is a real cost, we'd rather know it's worth it than guess.

- Approvals. Too much friction, or not enough? Would you want a per-agent "I trust this one, stop asking" switch, or does the pause serve a purpose?

- The planner. If you've thrown a genuinely multi-step request at an agent ("find the stalled deals, draft a follow-up for each, schedule them"), did it hold together or fall apart halfway?

- 24 templates. Which ones do you actually open? We suspect most people live in two or three and the rest is menu noise. Say which, and we'll stop pretending the others are equal.

- Kai and your local files. Comfortable, or a step too far?

- Visibility. We can now reconstruct any turn. Should that be exposed to you as a readable trace, or is that just more UI you'll never open?

What did you try, what did it do, what did you expect instead. Short is fine, a sentence about the thing that annoyed you is worth more than a feature request.

Thanks.