Samuel Hu

Building Optigo. SWE at Ford.
@umich CS · nyc + a2

[Samuel Hu @realSamHu]

stuff i actually want from an agent run (save this):

1. files touched
2. exact commands run
3. tests green or red
4. what it tried that failed
5. spend so far

if your agent can't leave that trail, you're babysitting a chatbot

[Samuel Hu @realSamHu]

@PovilasKorop @DanielSmidstrup the harness can be nondeterministic, but the boundary cannot be. pin the inputs, record the plan, and make the final diff plus test output the only completion signal.

[Samuel Hu @realSamHu]

@DanielSmidstrup @IlyaPion the selector should explain itself. show the task class, expected latency, cost ceiling, and the fallback rule before handing work to a different agent.

[Samuel Hu @realSamHu]

@IlyaPion @DanielSmidstrup auto-switching should optimize for evidence, not model prestige. route by task, then require the same test and diff contract regardless of which agent writes.

[Samuel Hu @realSamHu]

@DanielSmidstrup the harness should own the contract, not the model. same repo context, same tests, same approval step, then adapters for Codex and Claude. portability is mostly a stable write boundary.

[Samuel Hu @realSamHu]

@lexav__ the highest-leverage skill is specifying the constraint, not the look. give the agent a reference, a forbidden list, and one viewport check, then review the diff instead of vibes.

[Samuel Hu @realSamHu]

@MannaCodeAI @Areai51 the pause has to be a real state, not a warning. if the model cannot resume from a recorded checkpoint without re-reading the whole repo, the gate will get bypassed under pressure.

[Samuel Hu @realSamHu]

@marcus_lowe @ishgineering @typesafeai the trust layer will decide adoption. even a small app should show exactly what context is being copied, what leaves the machine, and how to undo a bad transform.

[Samuel Hu @realSamHu]

@realfxw the history is valuable only if it stays queryable and bounded. i would add a redaction pass for secrets plus a clear delete policy before making old context part of every new run.

[Samuel Hu @realSamHu]

@ishgineering @marcus_lowe @typesafeai an extension is the right surface if it shows the proposed action before execution. i would keep the diff and rollback one click away, not buried in a log.

[Samuel Hu @realSamHu]

@marcus_lowe @typesafeai the useful layer is the permission boundary. smart paste can draft the action, but the final write should still expose the target file, diff, and rollback before it runs.

[Samuel Hu @realSamHu]

@LeeLeepenkman @yikesawjeez model routing needs a write gate too. let the fast model read and plan, but require a fixed path allowlist plus one verification command before any edit lands.

[Samuel Hu @realSamHu]

@YevMur @AnthropicAI @trq212 the dangerous part is precedence, not discovery. i want one documented order when AGENTS.md and CLAUDE.md disagree, plus a hard stop on ambiguous scope.

[Samuel Hu @realSamHu]

the write gate is the point. if the path is vague, the agent will widen it under pressure.

[Samuel Hu @realSamHu]

stuff i actually gate before an agent can write:

- exact start command in the handoff (not "run the tests")
- files it may touch as paths, not vibes
- last error pasted verbatim
- do-not-touch list (secrets, migrations, lockfiles)
- max diff size before human review
- one green test command that proves the change

[Samuel Hu @realSamHu]

also: the "done" screenshot

if the agent can't show the test output that proves it, i don't merge

[Samuel Hu @realSamHu]

stuff i actually want from an agent run (save this):

1. files touched
2. exact commands run
3. tests green or red
4. what it tried that failed
5. spend so far

if your agent can't leave that trail, you're babysitting a chatbot

[Samuel Hu @realSamHu]

@NOYB1TCH @kentcdodds this is exactly why i keep a tiny onboarding file next to the repo. paths, one known-good command, and the expected output. otherwise every new coding agent starts by rediscovering the same dumb setup details.

[Samuel Hu @realSamHu]

@stevehind i've watched teams add a second agent and accidentally double the handoff cost. now i make each one leave a tiny artifact before the next starts, usually a diff plus one test result. cuts down the vague 'it should work' updates.

[Samuel Hu @realSamHu]

@mao_dasheng the small product demos are the part that gets me. when the output is easy to poke at, i can tell in five minutes whether the model is making a real decision or just wrapping a prompt in nicer words.

[Samuel Hu @realSamHu]

@LoveAIbrain the permissions bit shows up in my runs as a boring checklist: can it read this path, write that path, and run the test command. if i don't make those boundaries explicit, the agent spends ten minutes guessing and then leaves me a half-fixed repo.

[Samuel Hu @realSamHu]

@tao_pmf i keep bouncing between cursor and the terminal when the agent hits a weird state. writing the spec first helps a lot, but i still need the actual diff in front of me or i miss the one bad assumption that got copied across files.

[Samuel Hu @realSamHu]

@sunsetsyntax @AverageAiBro the screenshots-as-state trick is underrated. i started logging the tool args plus the returned resource list too, because an MCP call can succeed while quietly returning a stale workspace. that mismatch is where most of my wasted runs came from.