When harnesses compact conversations, skills fade.
Worse: echoes of their existence actively harm subsequent recall.
It turns out harness compaction choices and skill loading instructions combined make or break whether skills are lost or refreshed.
samsaccone.com/posts/when-me…
Your AI coding agent isn't slowed down by inference time alone, but by harness assumptions that treat process exit as the prerequisite for tool output consumption
In samsaccone.com/posts/harness… I explore Time to Awareness (TTA) and a path forward to unblock harnesses running human-centric CLIs to accelerate the agent iteration loop.
Sam Saccone retweeted
Imagine the amount of Java crashes this poor LLM must have in its training set that it decided 'NullPointerException' deserves its own token
Work no longer being proof of work is reshaping what it means to learn and what is value.
Verbiage around "slop" fundamentally stems from the new divide - The work to verify now often greatly exceeds the work to produce.
What once was a teaching moment becomes a meat proxy experience.
loved this idea - gerrit-review.googlesource.c… - turns out you can just ship 🚀
Establishing a "test-bench" application for your code is one of the highest ROI activities you can do on your project.
While it makes your manual testing easier, it unlocks your agents to actually verify without token-maxing on the verification struggle bus
Sam Saccone retweeted
Why not both? tracejam.xyz
🤔 Something transitioned with skills, computer use, models, and harnesses in the last 1 month making this now possible to run for 24+ hours on full operating systems with no intervention on my end.
Step 1. Spawn subagents to analyze the last 10k of resolved bugs and extract patterns of defects into a dossier
Step 2. Spawn investigation sub-agents leveraging the document to explore for defects looking at both source code and video interactions
Step 3. Spawn deep defect resolution agents to address the issue - record after videos - surface the fix and before after videos for manual review.
are y'all out here doing this but no one is really talking about it except @jarredsumner ?
You know, I always figured that the downfall of human society was going to be somehow coordinated inside of a phpBB forum
Madhu describes in a tweet what modern developer (now agent) productivity teams are doing every day.
About a year ago as agents gained reasonable tool following abilities it became clear "dev experience" became more about agent exp than human driven experience.
To do this well and at scale requires spending every day in the trenches of trajectories, evaling at scale, and working with harness and model teams to hill climb against your battle hardened benches...
(And yes I'm hiring 😅)
The best way to get good at evals is to take a workflow you know really well and figure out how to make its quality measurable.
Study the actual traces - the sequence of prompts typical users have, what good responses would look like at each step and for the end to end outcome.
Study where your product fails and create traces that capture them - messy tool call responses, missing context etc.
Once you have a good eval, think about how you would make it easy to run repeatedly and automatically.
Once you have that, think about how you will ensure your eval continues to mirror live traffic as your users patterns evolve.
This could be a series. Would you guys like me to write about it everyday?
. @fat how long until diffshub[dot]com supports skins.webamp.org/ - I need Doritos to truly unlock the vibes.
