In our Vercel AI SDK replays, agents learned the same build rule across 59 sessions and 40 tickets. One Codex run traced a missing import to a dependency's missing compiled output. What knowledge should you save for the next agent? dosu.dev/blog/coding-...

How do your Claude Code and Codex tokens break down? ( hint: you can find out with npx @dosu/decant )

Dosu@dosu.dev · last wk.

We gave Claude Code (Opus 5.5) and Codex (GPT 6.1 Sol) the same 49 Vercel/AI tickets, and we found some interesting gaps. Check out the full story (with graphs)! go.dosu.dev/agent-cost-g...

Does your coding agent choose a skill from the task...or only when you name the skill? Try a fresh session without the conversation where you wrote the instructions. Our procedural memory guide explains what to check. dosu.dev/blog/procedu...

After internal tests, we saw three embedding models rank undated Celery and DBOS facts in the same order for June and December questions. Your coding agent needs the history behind a fact to judge when to apply it. dosu.dev/blog/semanti...

Okay, your agent saved the instruction. Does it still make sense now that the code has changed? I would love to go back and see why we ended up there in the first place!

Dosu@dosu.dev · 2w ago

Your agent runs pytest again. In the last session, it learned that make test-int starts the database the tests need. That lesson may still be saved, but never retrieved. Before correcting your agent again, check what made it into context. go.dosu.dev/episodic-mem...

i told opus to remix the pdoom video and take liberties

I'd enjoy reading more research with a small example I could change as I go. This essay's argument for open software appeals to me because rerunning a result and exploring how it can change your understanding of the change. jepedersen.dk/blog/202505_...

It's Dosu Drop Day! 🎒 For our back-to-school edition, Dosu does the studying. When a coding session ends, Dosu writes down what your team would otherwise rediscover. 📝 Session-end study 🧾 Knowledge read cards 💳 Credits that go further go.dosu.dev/drop-sep26-bsky

The team keeps working after the agent’s session ends. @devstein.bsky.social on how merged PRs and resolved conversations keep agent memory current. Recording + transcript: dosu.dev/webinars/age...

Louise Deason suggests asking about the last promotion on a team before you join. I’ve also found it helpful to hear what that person worked on and how their role changed, so I can picture what growth might look like while working with a team. louisedeason.substack.com/p/nine-quest...

Nine Questions I Now Ask in Interviews That I Wish I'd Asked Five Years Ago

Or: how to interview the company while they think they're interviewing you.

louisedeason.substack.com

Marco Giancotti describes how reading Japanese fiction helped him feel familiar with everyday life in Japan. That makes me want to try reading novels set in a place I’m about to visit. aethermug.com/posts/backgr...

Backgrounding My Way to Japan

How to learn truth from fiction

aethermug.com

Astra, Sol, and Luna from OpenAI have new badges in Decant v0.6.0. ✨🔥🌙 See where your tokens went and what your agents did in each session, right on your machine. Your move, Anthropic! A themed Claude lineup would be fun to design around. github.com/dosu-ai/decant

Who reviews agent memory when people can’t keep up? Michael Mangus, Head of Product at Dosu, on trusting AI to maintain memory—and making mistakes easy to correct. Recording + transcript: dosu.dev/webinars/age...

Charity Majors describes asking for someone's opinion and getting an AI-written answer. I'd be disappointed too. I want to hear what you think, even if you're still figuring it out. charity.wtf/p/confession...

Confessions of an Unrepentant Slop Snob

A framework for thinking about when AI involvement is additive or a violation

charity.wtf

John Gruber linked to Tyler Stalman's comparison of iPhone photos and Kodak Portra 400 film. I enjoyed looking at the photos without labels and figuring out which ones I prefered more. daringfireball.net/linked/2026/... www.youtube.com/watch?v=m6cD...

iPhone 18 Pro: A Photographer's Review

Tyler Stalman's camera review, shared by John Gruber.

youtube.com

A summary of a summary keeps getting further from the source. Michael Mangus, Head of Product at Dosu, explains how repeated summaries lose details in agent memory. Recording + transcript: dosu.dev/webinars/age...

A contributor might want to keep their fork on a host they run. I'd like that to be an ordinary way to contribute to an open source project. The tools trying to make it work are part of my new piece, Git with the times. onlydole.substack.com/p/git-with-t...

Nix's description of returning to rough notes appeals to me because choosing what to rewrite sounds like a way to discover what I think. I'm curious about what we learn during that selection, even when most of the notes go unused. www.startingfromnix.com/p/if-experie...

if experience is the goal, then you always win

how to reinvent yourself: ft. the Miles Davis method, Georg Lichtenberg's Waste Books, Cate Hall's You Can Just Do Things

startingfromnix.com

Wrong memory can be worse than no memory. @devstein.bsky.social explains why stale agent memory can undo the value of remembering anything at all. dosu.dev/webinars/age...

I’m curious about using more open-weight models alongside frontier ones for daily work. Broad benchmarks don’t tell me much about how that would work in practice, so evals built around everyday tasks have been catching my attention.

Warp reliably restores the original task configuration for accurate comparisons. Simply swap the model, replay the task, and examine the proposed change. www.warp.dev/factories/be...

A fun critical thinking exercise for you or your team is to make a diagnosis before reviewing an agent's solution. If you disagree, the code review should specify the reasons why. That's helpful debugging context we often miss when starting with a proposed fix. spectrum.ieee.org/ai-engineer-...

A virtual cell is useful if it helps you choose the next lab experiment. That's the question I wanted to explore with Alexa Griffith in Attention Deficit Ep. 7. We also get into AI search and agents that won't stop. attentiondeficitpod.substack.com/p/attention-...

Can your agents identify which constraints you need to revisit before changing a design? entire.io/blog/introdu...

A code-search answer needs to let me open the file at the right line. I'm going to check the result, and hunting down the source of a summary adds work. turbopuffer.com/blog/large-s...

I love seeing coding agents make Blender scenes! Also, excellent scarf. bsky.app/profile/simo...

Simon Willison (@simonwillison.net)

New TIL on using Blender with coding agents on macOS: https://til.simonwillison.net/llms/blender-coding-agents-macos GPT-6 Astra (medium): > Use the already install /Applications/Blender to render a…

bsky.app

You don't have to fork an agent to make it useful for your project. Goose lets you change models and add MCP tools first. I wrote a blog about where to start, when a bigger change makes sense, and you should totally check it out! buff.ly/AE49AGm

What you can change when your agent is open source

A tour of Goose, from adding MCP tools to changing how the agent works.

onlydole.substack.com

Ever wonder what happens when you sign in to Netflix on your TV? You get shown a code, approve it on your phone, and never have to type in a password. Coding agents don't get that luxury, though. Learn about how we solved that problem in Dosu's CLI, and where we found our inspiration!

How to design a CLI login flow for coding agents

Design a CLI login for coding agents with a ticket flow that pauses, resumes in a new process, and returns the next command in JSON. Learn how Dosu built it.

go.dosu.dev

Alexa burned $3,000 in tokens in a day. I ran sixteen adversarial reviewers on a blog post. New Attention Deficit episode is out, where we talk about what agents cost, and when it makes sense (or cents) on when to give your agents more work. bit.ly/4d4EDJ5

Attention Deficit Ep. 6 – Sixteen Adversarial Reviewers

A $3,000 token day, Claude gets $534 to trade, Anthropic marks its text, and a Bash harness that knows when it’s time to go to sleep.

attentiondeficitpod.substack.com

Claude Code moved about 198 context tokens through its window for every output token in our PR replay study. Codex moved 134. This PR replay measures token flow and modeled API-rate spend. Check out the full breakdown in our latest blog go.dosu.dev/agent-budget...

Your coding agent budget pays for context, not code

Planning and gathering context account for two thirds of agent costs.

go.dosu.dev