olympian. crafting agents who craft agents
Joined March 2014
- Tweets 8,379
- Following 917
- Followers 6,792
- Likes 40,896
842 Photos and videos
Pinned Tweet
13 Sep 2025
The last time I generated an image in ChatGPT, it took 95 seconds.
It’s 2025. I didn’t want to wait that long.
That’s why I built @imageosai
It's
⚡ 6x faster than ChatGPT
🖼 Model agnostic (includes: Nano Banana, Seedream)
💸 100% free (for now).
download and lmk how you go
2
1,489
Jun 12
I’ve spent thousands of hours working with AI agents across a lot of different setups.
One problem keeps coming up: Where do you put the stuff you want your agents to remember?
Skills, markdown docs, transcripts, reusable code, images, bookmarks, tweets, YouTube videos, articles, notes, project context.
Sure you can throw it into Obsidian, Notion, GitHub, or a vector database.
But most of those come with setup, structure, or a learning curve.
So I’m building something much simpler:
A lightweight memory bank for your AI agents.
- No database setup.
- No infrastructure.
- No complicated workflows.
Just an easy way to share useful data with whatever agent you’re using: ChatGPT, Claude, Claude Code, Codex, or anything else.
If you’ve ever done useful work with AI and then thought “where do I save this so I can use it again?”, this is probably for you.
I’m looking for a few early beta testers.
Drop a comment or send me a message and I’ll get in touch.
2
2
202
Jun 10
“safe for general use” got me 🤣
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
227
Jun 10
Using @browserbase to give AI live web data and monitor competitor pricing
youtube.com/watch?v=hGUMwi4_…
144
Jun 8
If your not using /loops and /goals that's FINE.
Just stop using AI as ONLY a search engine.
That is the first big unlock.
Start with Claude Code. Start with Codex.
Prompt those coding agents. Deploy something to learn what these harnesses can actually do.
Jun 7
Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore.
You should be designing loops that prompt your agents.
2
522
Jun 6
My biggest takeaway from the @firecrawl challenge was how agentic work feels different from prompting. You are not trying to write the perfect instruction once. You are designing a loop that can observe, decide, act, validate, and learn from what failed.
youtu.be/m1-cyrMUshM
1
213
Jun 6
.@conductor_build has become my daily driver for AI-driven software development.
Here are the keyboard shortcuts I’m using most to move faster:
Meta:
⌘ / > View all keyboard shortcuts
⌘ K > Jump to another project
⌘ Shift F > View workspace history
Workspaces:
⌘ N > Start a new workspace
⌘ L > Focus the chat pane
⌘ T > Open a new tab or terminal
⌘ W > Close the current pane, tab, or terminal
⌘ [ or ⌘ ] > Move left or right between tabs
⌘ Shift A > Archive the current workspace
⌘ 1, ⌘ 2, etc. > Switch between workspaces
⌘ O > Open local directory options
⌘ Shift B > Open a localhost preview
Views:
⌘ ⌥ B > Open the right sidebar terminal
⌘ B > Open or close the left sidebar
PRs:
⌘ Shift P > Create a PR
⌘ Shift M > Merge a PR
Queued Messages:
Return > Send a message
⌘ Return > Queue a message
Arrow keys > Navigate queued messages
Delete > Remove a queued message
2
367
Jun 6
I built a stock research assistant that plans the research, waits for approval, fans out eight data tasks, streams the results, retries failures, and still produces a report.
youtu.be/6uDHgpRNUCE
202
Jun 5
This might be the best time to be working with AI.
Big companies are subsidising token usage to win market share, the same way Uber made rides feel unbelievably cheap when it launched against taxis.
Costs may keep falling.
Demand will keep rising.
It is a Darwinistic race.
144
Jun 4
Stop using AI like a search engine.
Here's my quick ladder from luddite to greek god:
L1. Chat (claude.ai/chatgpt)
2. Claude Desktop
3. Codex
4. Cursor
5. CLI in Cursor
6. CLI in terminal
7. cmux
L8. Conductor
213
Jun 4
Token usage resets = marketing 🔥
Jun 4
Hi. Over the last 24 hours we had three separate small incidents that affected Codex reliability. Those are three too many and we are taking active steps for them to not reproduce.
I have reset usage limits for Codex across all paid plans. May the tokens flow again.
1
242
Jun 2
used @triggerdotdev to automate stock research using some of their built in primitives.
very fun project
youtu.be/6uDHgpRNUCE
1
222
Jun 2
been using Cmux for the last 2-3 weeks, but started looking into @conductor_build lately, why should I not switch? ...opensource seems to be the only benefit of cmux to me at the moment?
174
May 28
I just beat Firecrawl’s $1,000,000 agent challenge: a series of timed tasks designed to test whether you can actually work with AI at the next level.
60 questions in just 240 seconds.
May 18
We're spending $1,000,000 hiring agent orchestrators at @firecrawl.
Just solve all 60 problems in the our CTF challenge, and we’ll make it very worth your while.
ctf.firecrawl.dev
2
1
802
May 28
Firecrawl is building infrastructure for AI agents to interact with real web data. That matters because agents won’t use the internet the way humans do.
Humans open a browser, read pages, click around, and search manually.
Agents can work differently.
They can call APIs directly.
They can capture data.
They can build memory.
They can route tasks.
They can validate outputs.
They can retry, improve, and keep looping until they reach the desired state.
The challenge had three parts:
1. Orchestrate: coordinate multiple agents to solve algorithm problems under time pressure.
2. Explore: search through massive codebases and answer precise questions despite imperfect information and prompt injection attempts.
3. Build: generate, modify, validate, and submit working code against a test suite.
The key lesson?
These were not “prompting” problems.
They were systems problems.
The winning approach was not asking ChatGPT for the answer. It was building agentic loops around the task:
- Capture the data.
- Clean the input.
- Identify what is known and unknown.
- Route work to the right process.
- Generate a solution.
- Validate it.
- Learn from failure.
- Try again with better memory.
That is where AI becomes powerful.
Not when it gives you one answer.
But when it can operate inside a system that improves with every attempt.
This is the shift companies need to understand.
We are moving from human-paced workflows to compute-paced workflows.
And you do not need to be a developer to see the importance of that.
But you do need to understand how agents work, what they are good at, and how they can be designed to handle messy, changing, adversarial environments.
The future of AI is not better search.
It is orchestration, exploration, building, validation, and iteration at scale.
1
372
May 27
My #1 though around AI that I keep coming back to when speaking to business owners.
"AI is no longer just a better search engine"
The shift happening now is toward agent-based systems.
AI that can reason through a goal, use tools, iterate, remember context, loop through problems, and operate at a scale humans simply can’t match manually.
You don’t need to be a developer to benefit from this shift.
But you do need to understand how these systems work, what they’re good at, and where they can create leverage.
The real advantage is not just faster answers.
It’s the ability to keep working through a problem when information is incomplete, conditions are changing, or the environment is adversarial.
All businesses face this. No business has perfect information to everything that's going on in their market.
We are moving from a world where the bottleneck was our own time to one where the bottleneck is the compute we can access and the agentic systems we know how to build around it.
That changes how companies solve problems, no matter the industry.
AI can help us understand faster, intercept earlier, test more paths, and move toward a desired end state with a level of speed and persistence that was not possible before.
104
May 22
very fun challenge. i solved all 3 levels with an elo of 3543. spent too much on L1 but I learned a lot. would love more of these :)
May 18
We're spending $1,000,000 hiring agent orchestrators at @firecrawl.
Just solve all 60 problems in the our CTF challenge, and we’ll make it very worth your while.
ctf.firecrawl.dev
1
303
May 14
every non-dev that wants to get into AI assisted coding is going to love this
May 13
Install ntn, the Notion CLI.
It brings the entire Notion API to your terminal, plus everything you need to build and deploy Workers. Built for humans and coding agents alike.
Install with: curl -fsSL ntn.dev | bash
2
354
May 1
I just walked through my exact environment for shipping websites and apps in hours. The whole thing is two apps. Cursor to view my files. Claude Code to do the building.
check the video here: youtu.be/P0opyzSt998
1
315