Introducing Kent 2.0 - a new version of my coding agent. The most significant change: Builder has been renamed to Kent. Still for builders, but now Kent is your true coding accomplice.

We also got a new domain at kent.sh. Migration will be described in detail when you update your Brew tap or run the migration script: essentially, you’ll get a new command to run and a new location for all Kent data.

Performance Fixes

This release includes dozens of fixes and improvements for both model behavior and quality of life. The main focus is performance, because I currently have a session that has been running continuously for 250+ compactions. At first Kent couldn’t handle that session’s file size (30+GB), so I had to urgently rewrite a lot of things. Performance is now great.

Also, we have a new brand color palette - brighter and more vibrant, consistent across the entire product.

System Prompt Customization

You can now customize system prompts more precisely. The system prompt is split into sections, and you can pick only what you need.

For example, for most subagents, the sections about delegation or code quality - slanted toward my own philosophy - aren’t necessary. Now you can use different templates in the prompt, build your own based on recommendations that I dynamically update in the repository, and add your own pieces or remove what you don’t need.

What’s the point? Creating custom subagents. I now have new subagents for plan review, code review, and plan drafting - with different settings and different system prompts. Works great.

I also made improvements so the model correctly launches subagents on its own, communicates with them, and independently uses the roles you give it. More details in the docs. Here’s an example subagent:

[subagents.plan_reviewer]
model = "gpt-5.5"
thinking_level = "xhigh"
description = "Extra-smart, costly, slow subagent for adversarial review of plans, architectures, specs, and implementation strategies. Use when the plan needs a hard quality gate or a second look before implementation. Can take a while to complete, give it time. Cannot edit files. Use --continue for re-reviews."
system_prompt_file = "agents/plan_reviewer.md"

description is what the model will see in its context. You can specify when and how to use this subagent there - similar to skills and agents in Claude Code.

Terminal Rendering Fixes

Dozens of fixes for terminal rendering: repeated prompts, disappearing prompts, vanishing messages, messages accumulating in the buffer, interrupted sessions, rendering that suddenly stops, duplicates, flickering of markdown - all fixed.

I’ll be honest: I’ve been really disappointed with Go for TUI, so I’m rewriting Kent’s terminal client from scratch. For now I’ve put in fixes - I know there are still issues and things break sometimes, but at least it’s usable now. I’ll try to ship the new terminal UI as soon as possible to put all these problems to rest.

New Slash Command /questions

The new /questions command lets you toggle the model’s ability to ask you questions. This doesn’t invalidate the cache - the tool simply fails with an error. It’s primarily useful for sessions where you want the model to temporarily leave you alone.

For example, if you’re going to sleep and don’t want a randomly asked question to pause work for the whole night. This has happened to me more than once during long-running tasks - so I finally added this command to turn questions off at night and back on in the morning. Very handy.

Goal Improvements

A lot of fixes for the goals feature:

  • The model now receives a reminder of its goal when starting a new session after a compact.
  • The model can no longer overwrite goals with random ones while continuing to work on the previous one. This was one of the causes of hallucinations.

System Sleep Prevention

Another important thing for long-running tasks - your laptop or computer going to sleep. My Mac kept falling asleep and work would stall. Implemented a config option that prevents system sleep. It’s enabled by default, but only when an agent is actively doing something.

So your Mac will smartly go to sleep only when no agents are running. More details in the docs.

Compact Fixes

Thanks to everyone who sent bug reports. Fixed issues with compact: sessions no longer just die when the context window is exceeded by a lot. Such recovery invalidates the cache - yes, annoying, but there’s no other way - but the session will now continue successfully after trimming the excess context that was blocking requests from being sent to the servers.

Local Models

Finally got the new Mac, so now I can dogfood local models. Made a bunch of fixes for them to work correctly and simplified prompting for local models to reduce hallucinations. A context window of at least 100k tokens is still recommended.

Would love to hear what you think about the name Kent :) I got really lucky with the domain - kent.sh is a great short domain, the name ranks well in search, easy to top. Kent comes from Russian “jail jargon” - meaning “Bro”.