Building AI agents + infra. Working on product loops, evals, that ship AI systems and own distribution.askmystack.space San Francisco, CAJoined July 2024
The most requested feature is often a symptom, not a need.
People ask for exports because the dashboard doesn't answer their question.
People ask for integrations because one step is too manual, that is there feature significance and company value grows with product enhancement.
Disaster imagery "within hours" sounds fast until you see the queue.
Book the pass, wait for the downlink, then a human sorts the images.
Put a change-detection model on the satellite and two of those steps disappear.
Everyone says "add more agents" for hard tasks.
I tested it on Cortex: 1 agent vs [N] on the same [N] tasks.
More agents: [X]% pass, [Y]× the tokens.
One agent with a tighter tool contract: [Z]%.
In simpler terms
Coordination costs more than intelligence buys.
B.
@lennysan@joshwoodward The team knows first because they see the usage logs and workarounds every day. Leaders see the roadmap. The fix is putting raw user sessions in the leader's week, not summaries.
@rafalwilinski@runlayer Folks work because each one owns its own memory and job, so context stops bleeding across tasks. Have you hit a case where folks needed the same context and drifted?
@svpino Graph retrieval helps when relationships matter, but building the graph is the real cost. Do you rebuild edges on every data change or batch it?
@mattpocockuk Splitting one big skill into [your number] small ones made the agent pick the right one more often. Where do you draw the line on skill length? curious!~
@hwchase17 We push results into a queue the parent checks between steps instead of interrupting mid-thought. Interrupts broke the parent's plan more often than they helped. Do you let the parent ignore a finished subagent, or must it acknowledge it?
@omarsar0 The catch is your harness logs need outcomes attached, not just diffs. A change log without "did the task pass after this" is noise for training. How are they labeling which harness changes actually helped?
Every AI startup pitch has a demo video.
None of them show the Slack thread where a human quietly fixed the issue on the product.
That thread is the real product insight that gives bigger picture about the product.
Most harnesses put one gate after the tool returns.
By then the charge cleared, the file moved, the message left.
Authorize outbound. Redact inbound. Same loop, two different products.
AI stopped being something you ask and became something that keeps working after you close the laptop.
Now it's in the room with your whole team.
So the product isn't the agent anymore. It's the approval step.
Autonomy isn't more agents. It's redesigning the job so AI Native that agent can finish and human become the orchestrator.
If your workflow still needs a person at every fork, you didn't ship leverage. You shipped a faster checklist.
“Tradeoffs that only pay off in 10 years will usually be underpriced.”
Play the long game. When you think in decades rather than quarters, you’ll see that your potential customers have trajectories, not fixed sizes.
Paul Graham on making startups more powerful:
paulgraham.com/powerful.html
The next leap isn’t a bigger chat model. It’s a decision model mid-run: pass, fail, retry cheap, escalate when it matters.
Most stacks optimize the paragraph. The expensive part is the fork.
Evals that only score the ending are reviews. Evals that change the next step are control of Safe AI.
Shipping agents overnight is easy. Waking up to a trail that tells you what changed, what it cost, and what to roll back is the hard product.
The scarce layer isn’t another model. It’s ops that still works when you’re asleep, So I started building an open source Fulcrum
github.com/askmy-stack/Fu…
The scarce skill in robotics isn’t another smoother grasp. It’s catching the almost-drop, writing it into a reusable recovery, and running the next carton without a human reset.
If your training set only has perfect demos, you taught the arm the easy day. The hard day is the product.
In AI times, demos are cheap and systems are expensive.
A demo answers once. A product answers next week with the same permissions, memory, and eval you can trust.
If it only works in the chat where you built it, you didn’t ship a product. You shipped a moment.
903 Followers 8K Following🚀 Introducing Rizq token – The Future of Growth-Driven Digital Assets
We are excited to introduce Rizk token, a next-generation crypto token.
779K Followers 673 Followingentrepreneurship zealot, grounded technology possibilist, believer in the power of ideas, passionate about sustainability & impact; @Seahawks & @12s fan
9K Followers 2K FollowingHead of Drug Discovery & Development Research @AnthropicAI | previously CTO & Co-Founder @CoefficientBio (acq. by @AnthropicAI) | ex-@Genentech | @MIT @Penn
9K Followers 1K FollowingML, systems, and everything in between. Building @guardrails_ai. Previously founding eng @predibase, @Apple SPG, @driveai_, @IllinoisCS, @iitdelhi.
928K Followers 481 FollowingFather of three, Creator of Ruby on Rails + Omarchy, Co-owner & CTO of 37signals, Shopify director, NYT best-selling author, and Le Mans 24h class-winner.
485K Followers 9K FollowingFounder/CEO of Henry Intelligent Machines and Creator Buddy ($300k ARR). Building a 100 trillion dollar economic engine
AI use cases: https://t.co/ovTDKYMqRt
43K Followers 2K FollowingStaff Research Engineer at @GitHubNext. But actually a designer. Writes about design, dev, & anthropology.
Also on Bluesky 🦋 @maggieappleton.com