- 1,500+ PRs Later: Spotify’s Journey with Our Background Coding Agent (Honk, Part 1):
Spotify’s control loop agent harness uses a cli command that runs an agent
with a prompt that is pulled off a task queue. The runner is a pod in a k8s
cluster.
- Building Pi With Pi: The pi
genAI harness is having trouble keeping up with all the llm generated prs. The
quality of them doesn’t sound great either. :/
- The Coming Loop:
Describes the way some companies are writing code now exclusively with a
clanker in a loop. I guess people are still there reviewing code?
- Build yourself flowers:
On the important of doing the work, building the skills and experience while
making use of new tools as they come along.
- Game Day: Stress-testing our response systems and processes:
Incident.io plans simulated incident response scenarios and carries them out
with members of the team as a way to practice and identify process
improvements … without needing the system to meltdown as part of the
exercise. :)
- Quantization from the ground up:
Brilliant article about quantization - why we do it, how it’s done, what are
the tradeoffs.
- Mark Carney keeps giving Danielle Smith everything she wants — and depriving her of what she needs :
Doesn’t seem like a new oil pipeline is the right thing to do right now for
anybody who’d need to be involved.
- What those AI benchmark numbers mean:
Really dodgy ways of verifying how well an llm can “think”. They can’t. What
does work is when a test appears in training data allowing an llm to have the
answer in its data set during inference.
- Your Agent is a Distributed System (and fails like one):
How does an agent system handle crashes and lost system state, slowness, split
brain, overlapping agents on the same task, etc. These are old problems and
llms are just one more instance of stuff we’ve seen before. There should be
good stories around stuff like this.
- Creating a security agent with Cohere North and Wiz:
An example of llm automation around the wiz vulnerability scanner. Looks like:
“Read this critical finding. Investigate and verify it in our environment.
Create a ticket in linear with your findings, asset information, and dates.
Mark the issue as in_progress in wiz. Notify the team if the finding is
serious.”
