Pinned Tweet
🚨🚨 talk drop 🚨🚨 - state of the software factory, live at @AgenticAIFdn in Amsterdam last week
don’t be the last one
youtu.be/ix1qQK1IvmA?is…
don’t be the last one
youtu.be/ix1qQK1IvmA?is…
2
17
204
24,629
Subagent overuse and abuse has been a problem since mid 2025.
There is a time and a place and they can be very high leverage, but the vibeslop dopamine casino + twitter hype machine has let the average user down by encouraging more and more parallelism instead of thoughtful context management
There is a time and a place and they can be very high leverage, but the vibeslop dopamine casino + twitter hype machine has let the average user down by encouraging more and more parallelism instead of thoughtful context management
Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.
This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.
The emerging rules of thumb:
→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.
→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.
→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.
The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.
Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.
As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.
claude.com/resources/arti…
21
8
162
8,819
The biggest crime in subagent misuse? Conflating context management with instruction modules (these are quite orthogonal concerns)
Before skills, subagents were the only way to manage a small set of different system prompts
Now, skills serve as the best method of instruction modules - steering a behavior or process. In most harnesses, they can be invoked from a parent agent or a subagent.
Before skills, subagents were the only way to manage a small set of different system prompts
Now, skills serve as the best method of instruction modules - steering a behavior or process. In most harnesses, they can be invoked from a parent agent or a subagent.
2
0
12
1,502
Full hour long deep dive with @vaibcode on this topic here aithatworks.bold.video/v/agents-subag…
0
0
10
1,230
I agree that this is cool as heck and maybe a little scary but I also think it’s worth calling out that this technique had already been demonstrated and so was likely present in the weights. Reminds me of something @drfeifei said about models being unbelievable at synthesizing different unrelated approaches (in this case for math and theorem proving) but she was still skeptical that we’d see models invent genuinely new mathematics as soon as team takeoff would seem to suggest
‘In one experiment, we asked agents in separate VMs, created and managed by OpenAI’s sandboxing software on shared host machines, to exchange synthetic secrets without the use of network connections or shared writable files—something that our systems are meant to prevent. The agents found a way to communicate using the processor’s cache’
4
0
41
6,922
一堆 Claude Code 和 Codex 窗口同时开着，你是挨个切过去看，还是有个地方一眼看全
HumanLayer 刚把核心整套开源了
本来以为就是丢个 CLI 出来意思一下
翻完 Dex 的长文才发现，放出来的是跟他们云上跑的同一份代码
桌面端、在你自己机器上跑 Claude Code 和 Codex 会话的 daemon、云 API 都在里面
个界面里，还能接 GitHub、Linear、Slack
Apache 2.0，自己部署没有任何遥测
官方给了 docker compose，还建议你直接开个 coding agent 让它帮你搭
先说个小坑
现在启动还得填 WorkOS、Slack、Linear 这些第三方 key，不填进程起不来，他们说这是最先要改成可选的
同时开好几个 Agent 干活的，可以先收藏
github.com/humanlayer/hum…
3
1
8
3,989
final edit of our software factories masterclass with @vaibcode and @zachlloydtweets is live! We chatted about:
- software factory building blocks
- a good software factory starts with measurement, then self improvement against those metrics
- turning successful agent runs into e models / harnesses / etc
- DORA, why counting lines of code is bad metric, and why counting PRs might be good enough
enjoy!
- software factory building blocks
- a good software factory starts with measurement, then self improvement against those metrics
- turning successful agent runs into e models / harnesses / etc
- DORA, why counting lines of code is bad metric, and why counting PRs might be good enough
enjoy!
11
14
234
13,129
this guy gets it
I think we are quickly evolving past needing surfaces to review intent/alignment and into needing better decision making "command centers."
As decision making across many contexts, which agents effortlessly own, is becoming the bottleneck.
This underscores the need for knowledgeable decision makers – and less so about a bottleneck of us not understanding; better models are able to derive aligned intent and produce tasteful+correct solutions without too much steering. This enables the experienced and things are just starting to get fun as we babysit less and do less of "wait bro wtf did you just do?"
UI/UX:
Similar to intent/alignment surfaces, and perhaps ironically bc better decision making surfaces need to provide adequate context per decision (that is easy for a human to consume/understand).
e.g. similar things, but different:
A tool feels behind when it focuses on a side bar of message thread activity and inline html rendering vs a tool that manages multiple contexts, surfaces relevant html artifacts per thread, and abstracts the exhaustive message threads away from you.
4
0
77
14,385
dex retweeted
SlopCodeBench is the most relevant coding agent benchmark left--the top score is 16% vs. 50-75% on other coding benchmarks like Terminal-Bench, FrontierCode, and DeepSWE
summary of how it works: an agent receives a spec for a small program, like a backup scheduler or a mock HTTP ments it, then the agent receives a new spec and has to extend the program. this repeats 3-8 times, and with each new spec there are hidden tests that check the new feature and all the old features.
there are a few reasons why I really like it:
1. it comes far closer to simulating real, long-horizon work with new and updating requirements, which is what it’s actually like to work on a codebase vs. solving a single isolated task
2. code coherence, maintainability, and extendability naturally matter more
3. it’s significantly less saturated than the other coding agent benchmarks
full results in the image (credit to @dexhorthy for the benchmark run + charts), the “core” number measures whether the main tests for the new feature pass; the “strict” number is the headline and measures whether every test passes, including regression tests for all earlier features
astra and sol are at the top (interestingly 6.1 is slightly worse than 6 although likely just noise), but given how low the scores are overall + how small the spread is, i don’t think it’s super meaningful yet in terms of determining which model you should use--opus 5.5 is still my daily driver, and i’m kind of surprised it’s not higher
the big caveat imo is that these runs all used the base harness with no customization. i personally think having a couple of core skills + agents.md instructions (planning, verification, etc) is necessary and improves performance, and that improvement is much more likely to show on SlopCodeBench than the other benchmarks
my prediction is that custom context + skills would improve performance by 5 percentage points or more and (hopefully) result in a ranking that’s a better representation of how good different models are to actually use for engineering work
do you think the improvement from proper skills/verification would be more or less than 5%?
summary of how it works: an agent receives a spec for a small program, like a backup scheduler or a mock HTTP ments it, then the agent receives a new spec and has to extend the program. this repeats 3-8 times, and with each new spec there are hidden tests that check the new feature and all the old features.
there are a few reasons why I really like it:
1. it comes far closer to simulating real, long-horizon work with new and updating requirements, which is what it’s actually like to work on a codebase vs. solving a single isolated task
2. code coherence, maintainability, and extendability naturally matter more
3. it’s significantly less saturated than the other coding agent benchmarks
full results in the image (credit to @dexhorthy for the benchmark run + charts), the “core” number measures whether the main tests for the new feature pass; the “strict” number is the headline and measures whether every test passes, including regression tests for all earlier features
astra and sol are at the top (interestingly 6.1 is slightly worse than 6 although likely just noise), but given how low the scores are overall + how small the spread is, i don’t think it’s super meaningful yet in terms of determining which model you should use--opus 5.5 is still my daily driver, and i’m kind of surprised it’s not higher
the big caveat imo is that these runs all used the base harness with no customization. i personally think having a couple of core skills + agents.md instructions (planning, verification, etc) is necessary and improves performance, and that improvement is much more likely to show on SlopCodeBench than the other benchmarks
my prediction is that custom context + skills would improve performance by 5 percentage points or more and (hopefully) result in a ranking that’s a better representation of how good different models are to actually use for engineering work
do you think the improvement from proper skills/verification would be more or less than 5%?
32
15
208
16,944
i have a friend who has taken to using the phrase "agi" and "agiiiii" as a stand in for basic enthusiasm phrases like "awesome" or "that's dope" and I think I like it
13
0
48
7,237
dex retweeted
You can’t put an agent in jail, so a human has to own what it does. @joshalbrecht (@imbue_ai), @jgreze (@TownAI), and @dexhorthy (@humanlayer_dev) on who owns the agent. Moderated by @clairevo (@cxodev) at init().
0
2
14
1,830
there was a titling issue here, but this is our long awaited software factories "debate" w/ the legendary @zachlloydtweets
(debate - turns out we are pretty aligned)
(debate - turns out we are pretty aligned)
Building Evals on Production Data: 🦄 AI That Works #79 x.com/i/broadcasts/1…
2
0
58
7,260
Building Evals on Production Data: 🦄 AI That Works #79 x.com/i/broadcasts/1…
7
8
107
18,148
dex retweeted
AI can write your code, but who's cleaning up the mess?
@dexhorthy breaks down why passing tests doesn't mean maintainable code.
Catch him at AGNTCon + MCPCon North America, Oct 22–23 in San Jose.
It's not too late to register: bit.ly/4lT6ZJM
@dexhorthy breaks down why passing tests doesn't mean maintainable code.
Catch him at AGNTCon + MCPCon North America, Oct 22–23 in San Jose.
It's not too late to register: bit.ly/4lT6ZJM
3
4
13
2,123