Head of Product & Engineering @kurrent_io, author of Hands-On Domain-Driven Design with .NET Core. Creator of https://xitter.cf/t.co/EPIgbC0WDQ. Views are my own.
Norway
Joined August 2010
- Tweets7.9K
- Following392
- Followers3.3K
- Likes4.1K
Alexey Zimarev 🇳🇴 retweeted
It's Friday! Here is a fun view from capacitor.kurrent.io/ of all your coding agents, prs, issues and colleagues live. We can do this because KCap records every coding-agent session and unlocks multiplayer AI.
#opus5_5_video #no_promise_this_makes_it_in_the_product
The bottleneck seems to be less PR review and a lot more CI time. Agents almost never delete tests but very happy to add more. My old server rack is worth running again with cloud compute prices going up. Plus, they can warm the house!
A new session gets the diff but not why it looks like that.
The context window is working memory, not storage. A resume surface has to carry the decisions and the ruled-out paths, or the next session re-tries what already failed.
What does yours hold that a diff doesn't?
tip for @humanlayer_dev on context engineering across sessions, compaction, and handing off to another human who will carry a session or task forward:
For sessions losing context on compact / new session, you should
1. make sure all context/decisions are into a doc in the artifacts (design, research, etc)
2. you can use a `/rpi:iterate-*` skill, e.g. `/rpi:iterate-design-discussion` to tell the agent in a new session to load context from artifacts
2b. You can also just @-mention the artifacts that matter in a new session
But the core approach is "offload all context to artifacts so resume / handoff is really easy" - nothing that matters should ever live ONLY in your context window
As kcap has a remote agent execution capabilities, we used it to add a review flow feature. Ask for an independent review and choose a harness, it spins up a fresh agent, receives a prompt from the implementer, then does the review and reports findings back. Unlike a sub agent, it won’t die at the end. It keeps the context to do consecutive rounds until it reports no findings. I use it all the time for code and specs.
Can your agent reach out to the collective prior work of the team to find what's relevant and avoid re-exploring everything from scratch? Mine can, for real
Agent pipelines get pitched as factory lines: stamp a unit, ship, next one starts clean.
A codebase never resets. Every agent-shipped change is context for the next run. Good patterns get copied. So do bad ones.
Where does the factory metaphor break for agent code?
yes definitely and also when you make a part or a car out of a factory, the quality of any output does not impact quality of future outputs, but if you’re building a codebase, everything you “ship out the door” impacts your ability to/ agility to ship the next thing, or adapt what you’re shipping later
Alexey Zimarev 🇳🇴 retweeted
I've handed coding-agent work to a new session mid-workflow, then watched ruled-outs, redirects, don't-try-X fail to follow. Handoff utility is a requirement, not a vibes hope the next agent "just knows." What do you insist travels when the next agent picks up?
I don't really know how "local only" people deal with team knowledge. I was taking over AWS Marketplace stuff from a colleague who left the org. Two things would've made my job a lot harder: (1) previous work was done with Claude on a machine that's not mine (2) that session would've been evicted anyway. With kcap.ai I had no issues, just kept asking questions and got everything done within minutes using the prior art that probably took a lot of time and tokens to figure out.
That can be a good addition to the kcap eval catalogue kurrent.io/features#evals
Indeed this is what I personally noticed as engineering leader when we started adopting coding with agents. Collaboration shrank, teams were turning into groups of ICs, each "pairing" with Claude Code or Codex. Then I started building kcap and we begun dog fooding almost immediately. Context awareness of coding sessions increased dramatically. Everything is open by default, private is opt-in. kcap.ai
Alexey Zimarev 🇳🇴 retweeted
⏰ Reminder: Part 4 of our 6-part series is next Thursday.
If an auditor, regulator, or customer security team asked you to justify a change your agent made — could you answer from the record? Or would you be reconstructing the story after the fact?
In this session, Madhu Murty runs a live audit on a real change to refund authorization logic — answering the five questions an auditor would actually ask, live, using nothing but the record. Then he shows how to turn the gap it exposes into a repeatable, automated control.
📅 Thurs, Aug 13 · 9am PT
🎯 For engineering, compliance, risk, and platform leaders with an audit, examination, or customer security review on the horizon.
👉 Register → buff.ly/lC5TOp1
#AIAgents #AgenticAI #CodingAgents #AIGovernance #Kurrent #Capacitor
Alexey Zimarev 🇳🇴 retweeted
😳 Agents write code fast. Now you've got a PR backlog.
That's the part nobody warned you about. Coding agents didn't just speed up writing code. They sped up producing PRs, and review didn't get faster to match. Even a review agent hits the same wall a human would: every PR takes time to understand, because a diff alone doesn't explain itself. (And you don't want your agents guessing, because they will. Confidently.)
Capacitor gives reviewers the reasoning behind the diffs in each PR - the plan, the prompts, the tradeoffs, the 'why' - instead of making them reconstruct it from scratch every time. Review moves at the speed the reasoning is available, not the speed a human (or agent) can reverse-engineer it.
The bottleneck was never how fast you could write code. It's how fast you can trust it. Capacitor closes that gap — so the backlog stops growing faster than your team can review it.
👉️ Try Capacitor free → buff.ly/RKcWSlw
#AgenticCoding #CodingAgents #AgenticEngineering #ClaudeCode #Cursor #Codex #Pi
Alexey Zimarev 🇳🇴 retweeted
🤖 Your coding agent just hit a wall. What happens next?
Most teams switch tools and start over. New agent, new session, zero memory of what was already tried. The next 45 minutes get spent re-discovering the past.
Capacitor fixes that.
Every coding agent session - Claude Code, Codex, Gemini CLI, Cursor, and more - gets recorded in its own shared stream. When you switch agents, the next one knows exactly what already happened: the hypotheses that failed, the fixes that didn't work, the reasoning that got you this far.
It doesn't start at turn one. It picks up at turn 48.
Switching agents should make your team faster, not reset the clock.
👉️ Try Capacitor free → buff.ly/RKcWSlw
#AgenticCoding #CodingAgents #AgenticEngineering #ClaudeCode #Cursor #Codex
Fable is great but I still ask Codex to review what Claude Code+Fable produces. So, we added review flows to Capacitor, fully automated back and forth between Claude and Codex for independent reviews. Codex keeps its own context during multi-turn reviews. capacitor.kurrent.io/docs/us…
I got fed up that colleagues can't take over my Claude sessions, that people don't see how my sessions went and I can't see theirs. That I can't recall Claude session in Codex and get unstuck.
Get out of vendor lock. Allow collaboration. Get back to paid programming and get peers to know why things were done that way and not the other. A small experiment, 100% dog fooding, now open for private preview sign ups:
Kurrent Capacitor capacitor.kurrent.io. Free for now. GitHub only (no Enterprise yet). #ai #claude #codex
We are hiring an engineer for the Core Database team ats.rippling.com/en-GB/kurre… #csharp #dotnet #hiring
I wrote a blog post about new secondary indexes in @kurrent_io (ex @eventstore) work and why would you want to switch from category and event type streams to indexes instead kurrent.io/blog/secondary-in…