you should check out what we've been working on for ref!
I've found something similar but rather than having one long lived thread, i have a plan doc that's extracted.
compaction has gotten great but i find the default pattern of having a sharable doc nicer than some ephemeral context engineering in a single long lived agent
all hosted, works with cursor, devin, warp. working on workflows soon!
although I'm super curious what your agent does when you're tired π
@vinvan oh amazing haha i have to work something like that into my workflow.
we've been thinking about how to automate that attention prioritization for folks and that's a really nice approach!
gamified checklists are one of my least favorite onboarding gimmicks but maybe they're actually perfect now?
1) every app is teaching people fundamentally new ways of working. you NEED to walk people up the stairs to new workflows somehow
2) usage based billing means you can meaningfully reward actions as people step through the app. literal dollars and cents!
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: https://t.co/bwn0sximKZ
What's a loop?
I'm not sure what people like @bcherny, @steipete, @tadasayy mean exactly but here/s my best guess at some basic loops anyone can pick, written as little programs.
cursorDebugMode()
agent adds logs to the codebase
while (human says bug exists):
- agent observes logs
- agent adjust code and logs
agent cleans up logs
β bug fixed
refLongLivedPlans()
agent drafts plan
while(human or teammates has questions about plan)
- agents researches and updates plan
human approves plan
while(plan is not complete)
- orchestrator agent starts sub agents
- impl agents impl
- review agent reviews
- orchestrator agent updates plan base on progress
- if(human needs input) orchestrator asks human
β feature complete
Importantly, both of these are "human in the loop".
You get more throughput by minimizing the number human steps and more importantly minimizing human context rebuilding per step.
The hard part of multitasking as a human is rebuilding context. If you can reduce a debug loop to "do the same set action, click Yes or No" that's a lot easier than doing any log parsing yourself.
Similarly, if all your ideas for a large projects are pulled out into a plan doc you can share with your team, rebuilding context is just "read the doc". Which you're already good at from the RFC reviews you did before AI changed your entire job.
I guess the goal is remove all human in the loop but that seems pretty hard and I haven't done it myself yet. One day!
goldilocks of ai coding skill progression:
"how do I stay in flow?" -> too cold. you're watching an agent do its thing. try multiple agents.
"i'm the bottleneck" -> too hot. you closed the loop, 1B token plaque, unimaginable amounts of well-designed code that no one uses
"i spend all my time thinking and then ship some stuff" -> just right.
I have been loving design mode but i've had some issues with the shortcuts (or maybe just ideas on how they could interact better?)
one thing that happens is ββ§D to start giving design feedback but then realize this should be a new context window so ββ§N and then want ββ§D focus back but now the UI/focus is in an odd state. Might be nice to have there a shift/cmd+enter or something to send my ββ§D input to a new context?
I'm typically doing a ton of back to back polish changes. and I try multitask mode and it also doesn't quite feel right. I end up just having one long context that i keep queueing messages for which leads the first problem I mentioned
@neal_k_patel stoked to try this. many features in ref's initial version were powered by gemini-flash-2.0
since then Gemini has been moving away from that sweet spot of cost/speed/intelligence they hit with 2.0
(also props on the nice ai ad spot!)
Today, weβre launching @TownAI: the AI assistant that learns you.
Weβre coming out of beta with a $55M Series A led by @ARampell at @a16z, with participation from @KirstenGreen at @forerunnervc and continued support from @firstround, @altcap, and @conviction.
Right now, getting real value from AI means prompting, configuring, building workflows, managing agents.
We think thatβs backwards.
The future of AI is a companion that already knows you and how you work. Town connects across your inbox, calendar, Slack, docs, messages, and workflows to understand what you need, then starts doing the work with you.
Drafting. Scheduling. Project tracking. Follow-ups. Context gathering. Multi-step tasks. And it only acts when you say so.
All adapting to your voice, priorities, routines, and relationships over time.
Your Townie is the AI assistant you actually need.
Devs who insist on working locally are going to get absolutely smoked by those who work in the cloud.
I've seen this first hand - you just can't scale past 2-3 concurrent threads locally - you hit local dev env hell super fast (colliding ports, work tree madness, etc).
Compare that to working in the cloud where you can instantly spin up 10 concurrent threads.
Completely independent, isolated VMs with their own checkouts.
This is how I shipped 131 PRs in 7 days.
You might wonder why the heck I need to ship that many PRs - it's because our first real law firm customer is finding all the rough edges which have to be quickly sanded down.
Also, I think $31.54/PR is a fabulous deal.
@Ydj79@karrisaarinen@linear sounds like you have your own thing AND now linear as options but we've been working on this exact thing at https://t.co/UUZduzn85j would love your feedback :)