Skills Enthusiast

I build tools that make agents and skills better at Uber. Ex-snap, ex-microsoft, ex-flexport. UM alumnus. Tweets represent my own opinions

[Skills Enthusiast @JasonVarbedian]

@tobi @markrmcqueen Sounds like you are still paying less tax as a % then others on income

[Skills Enthusiast @JasonVarbedian]

@jamesacowling Looped

[Skills Enthusiast @JasonVarbedian]

WebMCP is going to change everything. Agentic web soon

[Skills Enthusiast @JasonVarbedian]

@benhylak Try using https://t.co/jgNBDiDG4L. Put something in your agents.md on when to bother you

[Skills Enthusiast @JasonVarbedian]

@jamwt This is why I use convex for my apps

[Skills Enthusiast @JasonVarbedian]

@jhleath That’s interesting. So if you aren’t keeping the data use big ram? What about making it resumable?

[Skills Enthusiast @JasonVarbedian]

Long-horizon agents keep losing to interactive sessions for me.

I haven’t had much success with Fable or GPT-5.6 Sol at high reasoning. They churn through tokens, take forever, and still get off the rails quickly. That’s with my own harness, loop engineering, and skills I’ve tested extensively.

I get more done staying in the session with Opus, or using 5.6 Luna/Terra at lower reasoning. The feedback loop is tighter and the throughput is much better.
Maybe long-horizon agents work when a team has a backlog of pre-approved tasks and people available to supervise them. For solo engineering, more reasoning and autonomy have not translated into more completed work for me.

[Skills Enthusiast @JasonVarbedian]

@vboykis Well said. Do you have a skill.md for this concept?

[Skills Enthusiast @JasonVarbedian]

@jamesacowling 100%. The number of turns from human to output or verification is multiplying slop per turn.
Turns-since-truth * harness-quality

[Skills Enthusiast @JasonVarbedian]

@RhysSullivan Is it more things horizontally or deep into the task? I find I am doing too many things at once and so I can’t guide the inputs as well

[Skills Enthusiast @JasonVarbedian]

@levie Why prefer trained models to tailored skills?

[Skills Enthusiast @JasonVarbedian]

@satyanadella The learning loop should outlast any model. Private evals encode what good means. Memory, traces, feedback, and decisions preserve how the company learns. A decoupled orchestration layer lets you swap providers without rebuilding the loop.

That outer system is the meta-harness.

[Skills Enthusiast @JasonVarbedian]

We're building systems that call Claude Code, Codex, Cursor, or Cline and judge results. That outer system is a meta-harness.

The distinction matters because the same skill can improve one model-and-harness setup and make another worse.

https://t.co/1SExlIMXG1

[Skills Enthusiast @JasonVarbedian]

@dillon_mulroy How do you discover improvements like this? Do you ask your agent to read the docs

[Skills Enthusiast @JasonVarbedian]

@shugarDadddy Here's my attempt with attribution to you https://t.co/0MyklCdXpb

[Skills Enthusiast @JasonVarbedian]

@steveruizok Great idea. You should try using traces in combo with these. Could even open linear tickets

[Skills Enthusiast @JasonVarbedian]

Lots of talk about version control needs with single devs landing 100s of diffs a week (via loop eng). Hard to justify switching for smaller shops but great to have serious competition to push version control vendors

[Thomas Dohmke @ashtom]

42 is the answer to everything. This is @EntireHQ’s answer to Git in the era of agents: fast, independent, distributed. Mirror your GitHub repos, let your agents clone and pull from the region(s) of your choice, and…we're open sourcing it. 🤖

[Skills Enthusiast @JasonVarbedian]

@jamonholmgren Have you tried setting up tracing for your agents and having another review the session?

[Skills Enthusiast @JasonVarbedian]

@cansar This is a great mental model. Do you think correct solution can be determined? The rest seem eval-able. Do you also have a break down of when to use static vs dynamic?