Peter Bell retweetledi
Every company should have a full stack team of 5 building a product who are banned from directly writing code; they must force the agents to do it. In 2 months they will be your most productive team.
English
Peter Bell
19.5K posts
Peter Bell
@PeterBell
Helping CTOs scale AI adoption. Hands-on agentic engineer. Author @OReillyMedia. CTO communities @GatherDev. Exec summits @Kubecon_ Taught DS @Columbia_Biz
New York, NY Katılım Mart 2007
2.2K Takip Edilen4.4K Takipçiler
@JohnThilen @_lopopolo Once you have recipe runner and automated processes to interview spec review code and test the workflows it’s not that much of a lift…
English
I've tried setting similar pipelines up manually (n8n, autogen, crewai, etc) but it takes so long compared to the value I can think of. Therefore I just ask Hermes to do some half-assed version in case it becomes useful. If so, the value becomes apparent, and I can spend tokens ln improving it.
English
Most people do not know that they can just point their agents at my writing, tweets, podcasts, and talks and improve the output of their agents by 100x
English
I agree - I just personally see it as a feature, not a bug :) I do not think it's the one true way. I think it is one of the possible approaches - statistically the odds are it'll be one of those "remember when we used to use a single line perl script to run the internet and thought it was cool", but it's working for me for now, so I'll take the alpha and the learnings and we'll see where the pattenrs end up
English
@PeterBell Yeah I agree with you about context windows etc I just think we can build tools to manage them without siloing agents into different personalities.
English
Everyone is bragging about how many parallel agent claw swarms they're running but I suspect the right answer is just one superagent with a fat stack of skills.
English
@JoshPurtell For me a harness is a set of tools to allow any given model to perform better in a given set of tasks than it would with just one sho t or few shot prompting and foundation lab memory,
English
One thing I've noticed about Harness discussion is there are some people who think it means *the cybernetics work you do to connect models to value* and others see it as a convention for terminal interaction tool shapes and there is some talking past one another btwn them
Ryan Lopopolo@_lopopolo
While alternative coding harnesses may have short term lift, they will be bitter lesson’d away. I am bearish on any harness that doesn’t come from the lab whose model you are using. You’re fighting against post-training. To put a finer point on this, you know how like, ioctls are like “huh that's weird but I guess whatever it's what we've got we can work with that”? It is exact the same with like, the particular JSON construction the Codex shell tool uses. The model used to mangle nested quotes in this monstrosity RPC all the time but now it does not and it does not matter that the API is bad because billions of failed invocations are used to train to the harness we have, not the harness we deserve.
English
I have a multi step set of recipes for research where I capture sources, extract various types of insights, map them to node(s) in a knowledge graph and then add and enrich sources (people, companies, publications). Ingest is both standard web browsing and playwright srcipts where I have subscriptions and I have recency bias on a per node basis and a set of automated recurring scripts as well as the ability to run ad hoc research queries. Also working up multi-dimensional adversarial review of all sources and proposed insights with rubrics to ensure the quality and accuracy of the insights. Also working on finding a canonical source for stuff that's widely reported.
Amazing what you can do in a day when you have composable recipes, thoughtful end points and a nice agent orchestrator...
English
@PeterBell @_lopopolo Originally i did that with notebooklm, but i didn't like clicking around. Now I'm moving towards scheduled tasks (setup by Hermes) to scrape the known sources and convert into skills or into reviews of skills.
English
No question, more of the legal compliance and tax accounting advice/implementation.
But magic wand? OFFER HEALTH INSURANCE FOR SOLOPRENEURS with families in the US and no employees. Needs to be comparable to group plans you get when employed or through Justworks or Rippling.
it's gonna be one of the biggest problems if we see a expansion of the entrepreneurial class. It's literally my biggest issue - I often consider moving family to Canada. It's ridiculous.
English
@thomasbsauer Kind words. Magic wand, what should Stripe solve for founders _after_ incorporation? Big or small ideas welcome.
English
It’s an excellent platform.
10/10 would recommend.
Patrick Collison@patrickc
Stripe Atlas just hit 100,000 all-time incorporations. Q1 2026 is +130% Y/Y.
English
@metahacker_ Not my experience. Running a real business and exect to be able to 10-20x my revenue targets over 2 years leveraging automations.
English
The entire AI productivity meme breaks down once you realize that being able to do fake work even faster doesn't really change anything
I've vibecoded I don't know how many apps and the net economic impact of those apps is practically zero. The personal AI product I use the most is my own vibecoded AI trading bot and trade tracker. Works great, but, again, it has net zero economic or productive impact
Same for all the AI videos and images I've made. Most of it is entertainment disguised as productivity
We already had an abstractive economy where most of us were just doing various degrees of bullshit. None of the things I've ever sold - courses, monthly subscriptions, per-seat software - had any ties to economic or productive reality. I could offer 90% discounts and still make a profit at times because, again, the work was completely detached from the production of it
And now AI adds another deep layer of abstraction on top and it's somehow productive that you can pump out articles and blog posts and cheap code even faster
We've sort of memed ourselves, ngl
English
Why would human comprehension be the bar?
Current SoTA models are writing mathematical proofs that make no sense to me but I am told are elegant and correct.
No reason to believe that in the near future they'd write code that was beyond my comprehension as well. Just need to make sure they can fix it as well!
English
Can it even be "perfect code" when humans don't understand it? 🤔
Vikoo@vikrambuilds
Worst dev nightmare in 2026: AI gives perfect code… but you have no idea how it works 😭
English
I'm still mostly SoTA - until you have a rich orchesstrator pumping 24x7 I think reducing token spend is premature optimization. Once you have it, there's a lot of value in auto evals to determine smallest simplest model for each step, but unless you can run stuff 24x7 (lots of imagination required) you're probably better off renting a rig at runpod vs buying NVIDIA applicances, macbook studios or a bunch of H100's... run in the cloud until you prove you have the volume for a 3-5 month capital ROI, then feel free to go wild!
English
People who have maxed out their MacBook pros to run local AI models:
Do you use it?
English
Thats why you need to agentify the entire SDLC. User research (real and synthetic), competitive research, review of operating data, hypothesis proposal, rubrics of experiment selection, PRD creation, ADR's for the architecture (along with standard patterns and design systems) decomposition, coding, reviews for functional completeness, performance, usability, quality, maintainability, elegance of naming, and then of course deployment and incident response.
If you have a sufficientyl rifch, well trained pipeline with enough quality gates, the outputs don't have to suck.
English
We can do orders of magnitude more with agents, but it turns out that building bug free reliable software still takes a huge amount of effort. Now that we can do more, we have to much even more effort into making the software reliable.
You can see these companies crank out more features, but rarely are they high quality. You can feel the jank after the honeymoon phase of the first 5 minutes of use.
English
People understand that LLMs aren't actually "thinking," right?
Drexel-Alvernon, AZ 🇺🇸 English
I feel like we may sit around the holofire one day in the future telling our disbelieving grandkids of the depravities of the token famine of the fall of 2026 - "there were people stealing their kids gaming rigs just to get a few tokens a second overnight on a small, heavily quantized model"...
English
As @NataliaZarina says "you can feel the lack of compute"
Omar Shahine@OmarShahine
What the heck is going on with Claude Code. It spends a lot of time telling me things are going to take hours days or weeks to do and wants to stop working a lot. "Or if you'd rather I sleep on it and pick up tomorrow with fresh eyes, that's fine too." Like why would I want you to sleep on it? You don't sleep.
English
@ChrisJBakke Seriously. I run CTO communities at gather.dev and I do NOT envy the folks with 100+ or 1k+ person orgs :(
English
“You can still make $30M TC but you don’t have to manage a team of 3,600 people” is a pretty great value prop for a lot of public co CTOs
Henry Shi@henrythe9ths
Something strange is happening in tech. CTOs of billion dollar companies are quitting to take IC roles at Anthropic. Workday CTO -> MTS (Mar 2026) You[.]com CTO -> MTS (Mar 2026) Instagram CTO -> MTS (Jan 2026) Box CTO -> MTS (Dec 2025) Super[.]com CTO -> MTS (July 2025) Adept AI CTO -> MTS (Jan 2025) The mission is that real.
English
I think that ANYONE who cares about LLM performance turns off foundation model memory. It is effectively a cache that you can't invalidate or see or control reads or writes. Can't think of anything more painful.
Main reason I'm building my own orchestrator is t(a) to control syntax for composable recipes and (b) to manage memory and context to do things like explicitly say that if something is written to system of record it shoudl never be remembers (because another agent might change it in the db.
English
I found out I'm a little unusual vs most people in AI – I don't use memory. Here's why:
Stephen Biernacki@sbiernacki
Really great read from @hammer_mt at @every on the benefits of turning off memory in your LLMs. every.to/also-true-for-…
English
@hammer_mt I think channels in slack serve the same context decomposition purpose as named (human or role based) agents but I still see a lot of folks with claws and hermes doing a really poor job of managing their context windows.
English
I saw a paper but disagreed with the methodology - I have found a material difference when I load expert context into an agent. Tell them they're an x, separate;ly run research and then extract insights about the key competencies for that role. Summarize with specific frameworks, tools and actionable advice, decompose into sub domains so you can have a lot of context but they only load a subset per session and i've found that (for example) my Head of Research does certain things way better (and worse) than my head of sales, community and engineering. Give it a shot...
It's not about "play Tony Robbins" - it's about "here is a personality, back story, JD and set of tools to make you an amazing coach"
English
@PeterBell Ok but role prompting has been shown not to improve task performance (only affects response style) and you can fit yourself to a different UX if it's more productive.
English
Peter Bell retweetledi
To get access to the latest Gemini models, including the most recent experimental release that's currently at the top of the @lmsysorg leaderboard, use @googleaistudio.
You can also turn on features like code execution; use extended context (2M+); and turn off safety settings:
San Francisco, CA 🇺🇸 English
@_lopopolo Would be a lot easier if I had a company level budget and could afford stuff like firehose access but my entire computer/API/enrichment/token/hosting/subscription budget is maybe $2k a month so I gotta be scrappy right now...
English
Meh - it's OK, I am almost done with my research pipeline and it's dope. It includes deep enrichment. of individuals, publications and companies as well as OSS projects, case studies and common patterns and anti-patterns. I'll send you your packet end of the month to see how much I missed :)
English