@morganlinton One direct report sounds a lot like running agents well. Set the direction, let others run the day-to-day, and spend your time where it needs you most.
@morganlinton@VulcanBench 23/23 at every effort level suggests the bench may be saturated. Low to Max moves 0.57 points, inside the noise. Time for harder tasks, or a new direction? Long multi-step tasks might separate them again.
@stretchcloud The 42x makes sense for gh, which models have seen everywhere. I'd rerun it on a tool released after training. Then the CLI has to discover too, through --help and docs.
Curious if the gap holds.
The highest-leverage AI coding work I do isn't prompting. It's deleting.
Five versions of a helper? The agent copies one at random. I pick the right one, then tell it: find every copy and call site. Make this the only one.
@neilkanakia Curious how AI works end to end.
Binding needs a licensed producer, claims need a licensed adjuster, per state. Is a human signing off, or approving what the agent prepped?
@icanvardar One of the worst offenders is near-duplicates.Five versions of one helper, and AI clones a random one.
Pick the keeper and have the agent
remove the rest and update every call site.
@karrisaarinen@linear Flags make undo cheap, but they don't stay free. Agents ship behind them fast and nobody removes them.
I'd want flag cleanup on the agent's list too, or the codebase fills with dead branches.
Most early hiring mistakes aren't about skill. They're about hiring someone who needs a clear path when the work IS the path. I look for people who can figure out what they don't know, move across the stack, and ship without waiting for a perfect ticket.
@zeeg The splits that still earn their keep for me are write vs approve, and context vs context.
Cute names and per-bot rules help the human stay oriented. Under the hood it's still one model with the right memory and permissions.
@cryptodavidw This matches how I'm building.
Huge fuzzy surface → agent.
Small closed set → code I own and can read.
The trap is treating every edge case as fuzzy when some of them should just become tests.
@JJdoesTech@morganlinton@VulcanBench Drives me crazy when agents add irrelevant notes. “You told me not to do something so I’m not doing it.” is not a comment I need in the repo.
Agree on the gap. Agents struggle to size the repo and to size the task in front of them. They almost never ask how deep I want to go.
For Routine: try a skill that strips agent instructions and tests, then run the same suite with and without them. I'd bet the model and effort picks shift hard once the repo isn't pre-chewed for agents.
@Alex_Vainshtein Interesting. That hasn’t been my experience. I frequently get a ‘no’ when I ask agents if a feature or a bug fix is ready to deploy. What models are you using?
Honestly I've only seen it once. And that was a context miss, not the agent
hollowing out the suite to go green. I'm careful on test diffs, but I'm not redesigning the whole loop around it.
Atomic isolation is a good habit. Treating every test edit like a spec change feels like overbuilding for an edge case.
@morganlinton@MartinSolev@tobi Completely agree. It’s a foundational shift. Seems to be moving fast from my perspective. I’ve seen AI adoption across non-tech employees grow rapidly in the last 6 months.
@thorstenball The work that makes it better for agents
also makes it better for the humans. That's been my experience.
First thing I did at Rainwalk was assess
the platform from an agent's point of view. Adapting the setup is usually the right move.