Taylor Dolezal retweeted
This coming Monday, we're bringing together three incredibly talented engineers and founders to dig into a question and theme for the night: Are we running too many agents?
Josh Bleecher Snyder (@exedev)
@josh_wills (@datologyai)
@beyang (@ampcode)
We're already at 240 RSVPs! That's so fantastic. :-) This meetup is focused on a technical audience, so come learn and make some new friends. There will be an open Q&A at the end to ask questions and share thoughts. luma.com/0qr6bxcw
1
3
26
1,475
Taylor Dolezal retweeted
This feels short-sighted. The goal shouldn't be to measure the productivity of your individual developers.
The goal should be to build a system (i.e software factory) that minimizes the need for developers to be in the loop (i.e maximizes agent autonomy).
What if your AI budget went to the engineers who use it best?
Blanket usage caps are a hiring freeze in disguise.
Introducing efficiency-based auto-approvals for your most productive engineers, plus Devin Coach to help everyone get there.
devin.ai/blog/efficiency-bas…
1
1
2
169
Taylor Dolezal retweeted
One way @AmpCode and orbs feel like the future: in a new thread, said "audit the code base for this old API usage, group the code into areas, create linear issues for each area, and spin up a thread to fix each one" and I have a whole swarm of threads just ... doing it.
4
4
48
4,290
In our @vercel AI SDK replays, agents learned the same build rule across 59 sessions and 40 issues. We traced the repeated discoveries via @anthropicAI's Claude Code and @openai's Codex logs. What do your agents keep relearning? go.dosu.dev/session-memory-x
ALT Dosu studies an open glowing grimoire in a vast, candlelit cathedral library while enchanted pages float toward a moonlit stained-glass window.
1
1
3
220
Taylor Dolezal retweeted
Lots of people asked about this and I don’t think it makes sense to publish as a built-in skill:
github.com/docwriter-org/pla…
It’s from @sh_reya — I also * highly recommend * her and @HamelHusain’s evals course
We have a skill called “plain writing” in our repo that is the single most used skill internally by a factor of 2
7
22
442
40,991
Taylor Dolezal retweeted
Looking for an open source leadership opportunity? Join me on the @todogroup steering committee! We’re writing a book to help organizations run open source programs and exploring how AI agents can help us automate tasks.
Apply by Oct 20:
todogroup.org/community/2027…
1
3
6
600
Taylor Dolezal retweeted
great paper. saturation w.r.t self-learning is really interesting to measure for benchmarks, but it is hard to interpolate to agents at companies. at companies the underlying "game" is constantly changing.
excited to see more research like this
🤖 How well do agents learn from experience?
We study agents that amortize experience into reusable artifacts and introduce Agent Plasticity to measure how efficiently they improve on held-out environments.
Top performer ≠ most efficient learner.
1
3
273
Taylor Dolezal retweeted
Anthropic released Claude Haiku 5.5 today, so we ran the 5.5 family through our coding agent token usage benchmark.
Haiku, Sonnet, and Opus each did 5 tasks, 3 times each, in Claude Code at its default medium effort.
Haiku 5.5 spends a significant amount of time on reasoning and planning, which is typical for how smaller models increase intelligence.
7
3
8
438
Taylor Dolezal retweeted
early results on Luna vs Jev vs Decisions API on a few internal classification tasks
tldr; they are very comparable on performance and speed
2
1
3
384
As you use @claudeai or @OpenAI, you can improve your outcomes by understanding how your tokens are used (e.g., whether they're for knowledge management, research, routine tasks, or more).
We replayed 49 @vercel AI PRs across two coding agents, with an estimated $44.00 on Claude Code (Opus 5.5) and $27.29 on Codex (GPT 6.1 Sol).
Where did the gap come from? Spoiler: Low- and mid-scope work came out about even...
go.dosu.dev/agent-cost-gap-x
4
259
Taylor Dolezal retweeted
💥 I'm extremely excited to keynote AGNTCon in San Jose in a few weeks! I'll be speaking alongside legends like @timoreilly, @dsp_, and @techgirl1908
My talk will be on the future of the Agentic Web. See you there!
events.linuxfoundation.org/a…
8
10
60
8,566
Taylor Dolezal retweeted
Just shipped new controls to help manage spam and slop in private vulnerability reporting:
• Daily limits on new reports per account
• Custom daily limits for your repo
• Allow lists for trusted reporters
2
9
33
4,001
A working test command becomes a reusable procedure when the next coding agent knows when to run it and how to confirm the suite ran. We walk through make test-int, shared repository instructions, and skills. dosu.dev/blog/procedural-mem…
1
1
107
Can't wait to start plexing like it's 2099!
We like Rex, our new terminal multiplexer, so much that we wanted to share it with you. We’re admitting testers from our waitlist starting today.
superlogical.com/updates/pub…
1
182
Does your coding agent still recommend a tool your team replaced? We found a Celery migration comment that survived the move to DBOS and the removal of Celery. Semantic memory needs to preserve when a fact no longer applies. dosu.dev/blog/semantic-memor…
2
6
11
1,688