engineering @ ShopBack; creator of ai-devkit.com
Stockholm, Sweden
- Tweets 2,465
- Following 488
- Followers 608
- Likes 1,675
Pinned Tweet
I'm running Claude Code, Codex, and Gemini CLI in parallel.
The worst part isn't the token cost. It's the tab switching.
AI DevKit's Agent Console fixes this. One view, all your agents, send messages without leaving, quick hotkey to focus when you need detail.
Works in iTerm2, Ghostty, your default macOS terminal. No new app.
npx ai-devkit@latest agent console
4
4
15
4,610
Spent sometime this weekend to work on ai-devkit Devin CLI integration with SWE-2, share more soon
I love seeing models come out that give us top-tier performance at a lower cost. We need more of this.
Congrats @cognition!
Also, Cognition is giving 50 free Devin Max Plans for anyone wanting to try it!
Reply with what you'd with free tokens. Winners notified in 24 hours.
14
At the end of the day, write good code and design good design
My hot take from chatting to @poteto is that we should use MORE abstractions in the AI age
You can use them (combined with harsh lint rules) to reduce the design space available to the agent and constrain them only to good decisions.
Combined with the fact that high-leverage abstractions let you do more with less code - so, more token efficient.
Plus, unwinding the damage from a bad abstraction is much cheaper with agents.
This runs counter to a lot of folks thinking that agents just want to read the raw code. They can, but they're not maximally efficient that way.
Be braver! Design abstractions.
21
mean while my @github account cannot update payment method because the system tells that I’m in China!!!, support ticket has no answers for 3 months
1
81
AI DevKit memory gives agents a small local knowledge store for exactly that. The memory skill tells agents to search it before deep work, scoped to the repo they're in, and every record keeps the evidence behind it. You can browse everything in a local, read-only dashboard.
This is not a bigger prompt. It is retrieval: agents pull what's relevant when they need it.
What you get:
- Less repeated context. Decisions don't need re-explaining every session.
- Fewer rediscovered mistakes. A failure found once is stored with its fix.
- Cross-agent reuse. Claude Code, Codex, and any agent that can run the CLI read the same store.
- Human inspectable. Every record is browsable, with its evidence.
Quick start
1. Set up AI DevKit in your project:
npx ai-devkit@latest setup
Or
npm i -g ai-devkit && ai-devkit setup
This installs AI DevKit's built-in skills, including the memory skill. The skill tells your agent how to use memory: search before non-trivial work, store what it verified, and update entries that went stale. You don't need to run the memory CLI yourself.
2. Work as usual. Your agent searches memory before deep work and stores verified learnings when the work is done.
3. Inspect what it stored in the local dashboard:
npx ai-devkit@latest plugin add @ai-devkit/memory-dashboard
npx ai-devkit@latest memory-dashboard
Optional: the same commands your agent runs, if you want to add or look something up by hand
npx ai-devkit@latest memory store \
--title "Checkout retries reuse the original idempotency key" \
--content "Retry capture with the original idempotency key; never mint a new key on retry. Evidence: design review decision record." \
--tags "payments,design" \
--scope repo:checkout-service
npx ai-devkit@latest memory search --query "payment retry decision" --scope repo:checkout-service
Tips
- Scopes are global, project:name, or repo:name. Use the narrowest one that fits.
- Store what you verified (a test, command output, an explicit decision), not transcripts or guesses.
- Memory lives in a local SQLite file on your machine. The dashboard binds to 127.0.0.1 and is read-only.
4
3
255
Handing work between AI coding agents? Don't pass the transcript. I use Jev to classify every message (decision, evidence, blocker, next step, discard) and build a compact handoff.
Detail below
I built a session compact for AI DevKit, using Jev to label every message, and code keeps only what the next agent needs. Tested on a real session 1.72M → 5.9K tokens.
2
3
117
In this era, being open is better than keeping everything closed
It's moves like this OpenAI feels to be pulling ahead of Anthropic.
Anthropic is closed everything: closed source Claude Code, no you cannot use other models with CC, and so on.
OpenAI: Codex is open source, you can use other models; oh let's make it dead simple to use GLM-5.3:
53
never see this low bar from Google, we are also affected by this incident today
Oh my god, Firebase’s SDK started to crash ALL iOS apps that were using it, for ALL sessions, just like that, no way for any of these apps to do anything…
How amateur is all of this from any SDK, but especially one from a company with as high of an engineering bar like Google…
39
Hoang Nguyen retweeted
Disagree
just as high level languages help us code better - there's less to think about, it will be the same for AI. Why reach for ASM when you can use Rust - there's so much more that can go wrong. By this logic AI will also forego testing because they have so much intelligence to just reason through it.
Rust is a good prompt compilation target for the moment, but so is C++. And soon assembler. Then microcode. Myopic to think we're going to stop the agentic drill bit until it reaches computing bedrock.
57
53
888
87,602
I built a session compact for AI DevKit, using Jev to label every message, and code keeps only what the next agent needs. Tested on a real session 1.72M → 5.9K tokens.
1
5
323
Check out the code if you want to learn more github.com/codeaholicguy/ai-…
1
51