a gent
San Francisco, CA
Joined December 2009
- Tweets150
- Following631
- Followers230
- Likes12.4K
A few months ago I thought I was buying a second monitor so I could monitor more agents. Turns out I prefer working on a single monitor and it's better to talk to fewer agents so now I have a 6k 34" doordash monitor.
Agentic development is like whack-a-mole. But the moles can create more moles.
I realised the majority of my agent session time was spent waiting on tests to run, so I made a distributed test runner on @modal, dropping the full suite wall time from 30 mins to 2 mins. But now I keep blowing through my weekly quotas in a few days per account.
So now it’s time to start analysing the agent sessions to figure out wtf are they doing. They do produce (mostly) accurate work, but depend on a lot of review/fix loops. Which previously were indirectly rate limited by slow test runs.
As the cost of performing work goes down, the importance of doing the right thing at any one moment is lessened if you can guide the trajectory of the work accurately. But it’s just pushing the bottleneck around to different places. The way you work has always been a spectrum of tradeoffs of time, accuracy, price, and sanity. Agentic development lets you bounce around different points on that spectrum and I don’t think the industry has found the right place yet.
if you spend a bit of time working with agents you start to see that every "agent first" company's open source agent products are the default outputs you get. Everything is either Foreman or Forge.
You can build your own software factory with eve.
Deploy the template and let Foreman assist you in building your project, closing GitHub issues, and more.
github.com/vercel-labs/eve-s…
I have sam's-house-chores bench which so far every frontier model has scored 0% on.
Open sourced below:
1) the right combination of settings to use a single tv remote with a ps5, nvidia shield, and pc connected to a philips hue sync box connected to the tv which also has a sound system connected via eArc. So far the best result is being able to do everything with one remote other than turning off the tv which requires another remote and switching hue inputs which I do on my phone since auto-switching works maybe 20% of the time.
2) the right combination of settings to allow me to have a single cable between my monitor and laptop handling data and power, and my half a million random items plugged into the monitor. This one actually worked for a week but then began to randomly stop providing power to a different random subset of devices each day. I have declared bankruptcy and now have endless cables. Life is pain.
One agent has been talking about the "test estate" recently, turns out rather than an agentism it's my transcription tool mishearing "test suite"
It’s been graph engineering all along.
A prompt is a node.
A thread is a path.
A subagent is a branch.
A loop is a cycle.
A multi-agent system is a graph of graphs, so, a graph.
What's the true core, category engineering?
My codebase greeting me in the morning after fable downgrades to opus sometime overnight, likely trying to pen test the treat drawer
jfc, I think "agentic" just means able to go on the technical version of a bender. Overnight fable got stuck fixing a node version issue (happens to us all), decided to build a custom sandbox thing, then decided to pen-test it, hit the "hacking" internal classifier, got lobotomized to opus and proceeded to produce absolute nonsense code until hitting a 5hr limit and the whole mess ended.
