Product & Research Lead. Token Junkie. Formerly of @twitter, @apple, @google. Proud husband to @kristineyjones π«
Joined July 2008
- Tweets 361
- Following 1,048
- Followers 8,895
- Likes 2,329
91 Photos and videos
Jun 13
The best Codex automations aren't code. They're five weekend routines that clear Monday before I get there. β
1
29
Jun 12
Mythos has been rumored forever, but is this just another Opus model upgrade, or did the ceiling for software work actually move with Claude Fable 5? π
2
25
Jun 11
Loop Engineering is the next leap: agents find work, execute, evaluate, and repeat without constant supervision.
The secret? Build the evaluation first. Then let the loops run publishing to queues with @linear as the broker.
3
125
Lots of people asked how I used Fable to edit its own launch video so I made a video about that!
TLDR it wrote a lot of code & tool calls to use transcription services, ffmpeg, do colorgrading, use the figma mcp, make remotion UI and render it.
I didn't touch a video editor.
293
620
9,235
978,160
Gareth Paul Jones π retweeted
Jun 10
Things I really dislike about Fable:
1. Anthropic collects my prompt history, stores it, and does whatever they want with it for 30 days. No opt-out
2. They can nerf their most expensive model without telling me, billing me the same amount, wasting my time. Whenever they want
197
335
7,000
415,316
Jun 9
I pushed 3.5k PRs yesterday and I think @mattpocockuk is right here. Loop into a queue then queue into a loop. Queues provide structure and a bunch of nice human readable traces to the loops.
The best solution I have is a custom Symphony that runs via @linear with the agents communicating and making state changes to issues and nested issues. The fleet runs on linear and then everything else gets distributed. Loops publish artifacts to linear and then are queued to run other loops. π π§ββοΈπ§ββοΈπ§ββοΈ
Jun 9
Everyone's banging on about loops
When they should be thinking about queues
3
99
Jun 8
If you loop engineer and run more than 10-20 Codex/Claude loops 24/7 on premium hardware today. youβll hit OOM issues. In 6 months everyone is going to be cap constrained by hardware and energy.
Good take
My guess is
- demand for intelligence is near infinite
- but 80% of workloads will be running on 99% cheaper models within 12-18 months
- 20% of workloads will still run on latest gen models where IQ maxing is important (scientific breakthroughs, higher level ochestrator agents?)
- rough analogy might be what % of macbooks or gaming PCs sold have the maxed out specs for CPU/GPU, prices are falling much faster than Moore's law here though
- this leads me to think the limiting factor will be energy and compute, not better models
At Coinbase we're working hard on routing prompts to cheaper models where appropriate, and in some cases have been able to keep costs roughly flat, while token usage continues to grow exponentially.
71
Gareth Paul Jones π retweeted
Hereβs your monthly reminder that you shouldnβt be prompting coding agents anymore.
You should be designing loops that prompt your agents.
1,786
1,376
19,584
8,308,920
May 16
AI coding tools are starting to feel like slot machines.
If you find your self saying "ohh that's almost right, let me try one more prompt again" loop, in this video i talk about how i'm trying to avoid casino mode and what i do instead.
4
128
May 9
Codex Orchestration w/Symphony is timemaxxing for builders.
It turns project work into isolated coding runs with agents that handle CI proof, PR reviews, complexity checks, walkthroughs, validation and safe-healing merges.
This might sound hyperbolic... but it could make folks way more productive. It removes a bunch friction with βagent made a PR slopβ, "wtf happened in this Github Actions", βhow to avoid doing human reviews on 100 PRs a day" and more.
In this video, i dig into my experience using it with where it helped and how to avoid hitting a rate limit on on 1000 merge conflicts in less than 24 hours.
github.com/openai/symphony/
3
106
May 3
The new /goal feature lets Codex create, pause, resume, and clear long-running objectives across sessions, with runtime continuation behind it. We dig into see how useful it is. x.com/i/broadcasts/1rGmqomkRβ¦
2
73
Apr 25
Agent management gets way better once you stop treating everything like one giant thread. Iβve been using βswimlanesβ separate vertical lanes and it makes management way less chaotic. x.com/i/broadcasts/1dJrPENnDβ¦
2
75
Apr 11
Most AI coding demos are βlook, it made an app.β
This one is different.
Itβs:
- hereβs the metric
- hereβs the sandbox
- hereβs the experiment budget
- improve it
- if you fail, reset
- if you win, commit
- never stop x.com/i/broadcasts/1oKMvRYVdβ¦
3
100
Gareth Paul Jones π retweeted
Itβs very unclear to me what the upper bound on daily token use per person is going to. Orders of magnitude beyond this for sure.
Magical OpenClaw experiences that use frontier models cost $300-1,000/day today, heading to $10,000/day and more. The future shape of the entire technology industry will be how to drive that to $20/month.
150
60
1,314
216,298
Apr 4
Building a personal agent with a fully local setup w/Hermes Agent Gemma 4.
Is Gemma 4 any good? How does Hermes compare to OpenClaw? How painful is self-hosting?
Only one way to find out. x.com/i/broadcasts/1yKAPMkzdβ¦
1
3
134
Mar 28
Codex plugins are pretty great. Bundling skills, workflows, and context into something portable across projects is powerful. Are they any good? x.com/i/broadcasts/1PKqrEjVZβ¦
3
87
Mar 21
The biggest bottleneck in AI coding might not be intelligence, it might be structural: one agent trying to do research, implementation, review, and cleanup inside a single polluted thread. If you use Claude Code this may be the case to make the switch. x.com/i/broadcasts/1RJjpzOjNβ¦
4
132
Gareth Paul Jones π retweeted
Mar 20
Even though every AI company is building their own version of OpenClaw (which is smart!), I haven't seen any of them get anywhere near the love and passion that OpenClaw inspires.
There's something special about the OpenClaw experience that's hard to copy.
We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord.
Use this to message Claude Code directly from your phone.
71
17
236
37,608
Mar 14
Claude Code doesn't just write code anymore. It can spawns entire agent teams and coordinates them autonomously. It's wild! x.com/i/broadcasts/1XxygmBqrβ¦
4
143