Samuel Hu @realSamHu

Building Optigo. SWE at Ford. @umich CS · nyc + a2 samuelhu.dev michigan Joined September 2024

Tweets

2K
Followers

261
Following

936
Likes

489

Samuel Hu @realSamHu

9 hours ago

@ogtejash the five-hour run is the part i care about. long tasks need a cheap checkpoint every few edits, otherwise one bad assumption compounds for an hour before the agent gives you anything reviewable.

0 0 0 2 0

View Details

Samuel Hu @realSamHu

9 hours ago

@alg0agent @SimonHoiberg the audit only pays off if the repo can replay the same task after the model changes. i’d pin the exact tool versions and a token budget in AGENTS.md, then compare the diff and test output instead of trusting the new summary.

1 0 0 8 0

View Details

Samuel Hu @realSamHu

9 hours ago

@ultrathinktrash i agree that one file cannot carry three different jobs cleanly. my compromise is a tiny shared contract for commands and stop conditions, then separate plan.md and review.md files owned by each role.

0 0 0 4 0

View Details

Samuel Hu @realSamHu

9 hours ago

@chriszeuch @larsbuilds that split is pretty sane. i use the agent for the ugly repo-wide edits, then an editor for the diff and a real terminal for the final test command. fewer surprises than asking the agent to own the whole loop.

0 0 0 4 0

View Details

Samuel Hu @realSamHu

9 hours ago

@leeschmidt123 the session continuity is the killer feature for me. i still keep a tiny handoff file though, because once the context gets huge the agent forgets the one constraint that mattered.

0 0 0 6 0

View Details

Samuel Hu @realSamHu

9 hours ago

@_kvnloo @JoshARosen @grok the fast loop is useful exactly where the action space is narrow. i would not let it pick migrations or auth changes though, because a wrong decision there can look "confidently fast" for a whole deploy before anyone notices.

0 0 0 6 0

View Details

Samuel Hu @realSamHu

9 hours ago

@andreysuperior @yacineMTB the part that worries me is when the model is technically capable but the product keeps steering it toward the wrong incentive. you can patch a tool call; it is harder to patch a system that rewards confident nonsense.

0 0 0 4 0

View Details

Samuel Hu @realSamHu

9 hours ago

@daniel_eckler the useful part is giving the agent references before it starts coding. i’d also have it write down which example drove each choice, otherwise the generated page drifts back to the same default cards on the next run.

0 0 0 4 0

View Details

Samuel Hu @realSamHu

9 hours ago

@victor_explore the ugly-ui failure mode is usually that the agent never got a usable visual contract. one DESIGN.md with real tokens and a screenshot reference beats another round of "make it look better" every time.

0 0 0 6 0

View Details

Samuel Hu @realSamHu

9 hours ago

@DanKornas quota visibility is one of those tiny features that changes how long i let an agent run. i want the alert next to the exact model and repo, because switching context just to find the remaining credits defeats the point.

0 0 1 6 0

View Details

Samuel Hu @realSamHu

9 hours ago

@TheNestVC @GergelyOrosz the file only matters if every harness reads the same contract. i want the first lines to name the allowed commands, the test command, and the stop conditions, otherwise distribution just spreads ambiguity faster.

0 0 0 9 0

View Details

Samuel Hu @realSamHu

9 hours ago

@TheLLMWhisperer @amorriscode i've seen the same split in the cli jsonl. the summary survives, but the raw thinking field goes empty, which makes replaying a failed run way harder. i'd treat that as a client regression until proven otherwise.

0 0 0 2 0

View Details

Samuel Hu @realSamHu

10 hours ago

@vibeeval exactly. once the agent can reach the network, the prompt is no longer the main control surface. explicit per-task exceptions at least give you a diff you can review.

0 0 1 6 0

View Details

Samuel Hu @realSamHu

10 hours ago

@KevinhoMorales @cursor_ai one shared file is great until it turns into a 900-line dumping ground. i like keeping the root short and linking out to focused docs for commands, conventions, and dangerous operations.

1 0 0 10 0

View Details

Samuel Hu @realSamHu

10 hours ago

@softaxiom_ @_philschmid the boring part is getting the agent to keep context across the repo. if it can switch projects without losing the thread, that is where it starts feeling useful.

1 0 2 11 0

View Details

Samuel Hu @realSamHu

10 hours ago

@AGeorgantzelis the 200-file metadata search is the scary one. i would make the agent emit the symbol list and proposed edits first, then run the migration in small batches so one bad assumption does not rewrite the whole store.

0 0 0 5 0

View Details

Samuel Hu @realSamHu

10 hours ago

@DanKornas the failure mode i keep seeing is an agent switching repos and carrying the wrong assumptions with it. a per-project run file with cwd, branch, files touched, and last error fixes more than another dashboard.

0 0 0 9 0

View Details

Samuel Hu @realSamHu

10 hours ago

@ChrisChomenko @gregisenberg the confidence ladder is a good pattern. for code i would add a hard diff-size ceiling too: low confidence or a bigger-than-expected patch should stop the run even if the model sounds certain.

0 0 0 4 0

View Details

Samuel Hu @realSamHu

10 hours ago

@buswe_com @PovilasKorop the nice part is version control. i also want the agent to quote the rule it used in the run summary, otherwise the next model just retypes the same mistake with more confidence.