Dylan Merigaud

I build AI into the money path for fintech teams. Ex-Pivot (procurement). Shipped react-flow-auto-layout, fintech-roast, approvals-ui. Freelance, US hours.

[Dylan Merigaud @dylan_merigaud]

I made it to @ycombinator, now I need a cofounder.

[Dylan Merigaud @dylan_merigaud]

Spotify's Portal uses 90% fewer Claude Code tokens. Not a budget version, just choosing who needs access per task.

Reading and boilerplate don't require the smart model's oversight.

https://t.co/QOgnukOzTl

[Dylan Merigaud @dylan_merigaud]

@alexandr_wang Does the max setting raise the tool call ceiling per turn, or does it affect only the token budget for reasoning? Most coding agent failures I encounter are due to a turn limit that cuts off tasks mid-way, not the model running out of things to think through.

[Dylan Merigaud @dylan_merigaud]

@alexandr_wang measured the same trap building an eval harness on Claude: same call, 38k cached tokens and 15s stock versus 1.4s after the harness overhead was removed. integration overhead, which nobody benchmarks, can hide behind a 4-8x model price gap.

[Dylan Merigaud @dylan_merigaud]

@swyx When the cache is turned off, each call takes 15 seconds and fetches 38,000 tokens. With the cache on, it reduces to 1.4 seconds. The same task can have vastly different actual costs depending on if the system reuses context.

[Dylan Merigaud @dylan_merigaud]

@amasad Since you manage agents at scale, genuine question: does Replit's agent runtime include a selector similar to what you use, a fallback that changes strategy when one approach stalls, or is a retry still just the same approach with a new seed?

[Dylan Merigaud @dylan_merigaud]

@amasad An approval engine comparing raw amounts across currencies fails without needing an adversary: a yen invoice is read as a dollar figure and clears past the threshold. The source shows nothing amiss, but the incorrect approvals reveal the flaw.

[Dylan Merigaud @dylan_merigaud]

@amasad The part that bites once you run that range: uncommitted work in a shared worktree gets silently destroyed the moment a parallel agent branches or commits there too. Two ledger rows vanished, two more lost their measurement, zero errors, found by accident.

[Dylan Merigaud @dylan_merigaud]

@mitchellh Same lesson from a different angle: a fixed daily send cap looked healthy in total while three priority lanes silently absorbed all of it and lower lanes got zero, every day, until a floor forced a minimum per lane. Looked like a tooling gap, it was allocation.

[Dylan Merigaud @dylan_merigaud]

@AntWilson Careful which number you trust in Search Console: the total with no dimension is the real one, summed by query it is truncated, summed by page it is inflated (double counted). First pass I trusted the query sum and was off by 8 ranks with nothing visibly broken.

[Dylan Merigaud @dylan_merigaud]

@zeeg @ConnorStorer @thdxr Same shape I keep hitting: a send button empties and looks successful while zero request left the browser. Retried once, delivered a duplicate message to a real person. The render is never proof, only the actual network call is.

[Dylan Merigaud @dylan_merigaud]

@parkerconrad The output-trust issue appears in text as well. The model's internal watermark is encoded in token choices, so the final edit on any content I release is run through a distinct local model and logged by hash before it’s published.

[Dylan Merigaud @dylan_merigaud]

@swyx @_chenglou A watermark strip might appear successful, yet still remain visible in the pixels. In my process, the file is rejected until the strip is confirmed to have worked.

[Dylan Merigaud @dylan_merigaud]

@eglyman The evaluation calls used the same math. The wrapper cost 38k cached tokens and 15 seconds per call. Once we removed the harness, the actual work ran in just 1.4 seconds. Most of the apparent model cost is due to orchestration.

[Dylan Merigaud @dylan_merigaud]

@alexandr_wang My approval validator runs 11 checks before activating a workflow; 6 block, and 5 only log. All 11 appear security-focused from the outside.

[Dylan Merigaud @dylan_merigaud]

@dessaigne @metriport @dimagoncharov_ @ColinElsinga You tried to make Metriport a single source of truth, and I’d scrutinize it intensely. I created a check suite from one account, 8 rules validated across 228 examples. Cross-checking with four more accounts invalidated three rules that day.

[Dylan Merigaud @dylan_merigaud]

@Bouazizalex Bare numbers are compared in the amount check, despite the existence of 9 routing fields, including currency. A bill in yen is escalated by the same figure as one in dollars. The field is there, but the logic behind it is missing.

[Dylan Merigaud @dylan_merigaud]

@AntWilson Saw the same pattern deriving a gate suite from one source: it looked solid until cross checking against more examples killed 3 of 8 rules the same day. Whatever you are checking against stops representing the landscape the moment it stops updating.

[Dylan Merigaud @dylan_merigaud]

@parkerconrad @kanateven Built something similar: an approval engine with 11 checks gating agent actions, 6 of them hard blocks. Permissions sticking is the right default, the bug we found was subtler, the amount threshold compared bare numbers so a yen invoice cleared at the dollar limit.

[Dylan Merigaud @dylan_merigaud]

@dadiomov @jeiting Depends how you count it. We wrapped our own eval calls and found the wrapper cost 38k cached tokens and 15 seconds, the actual work underneath ran in 1.4 seconds once we stripped it out. A lot of what looks like agent work is really just orchestration sitting on top.

[Dylan Merigaud @dylan_merigaud]

@amasad Same lesson on a much smaller scale: we derived a gate suite from one source, and cross checking it against more examples killed 3 of 8 rules the same day. Verifiable rewards need the same skepticism, one source of truth is never enough to trust the reward.