Adam · Arcyton

Founder @Arcyton — venture studio for data infrastructure, scrapers & AI. 50+ production actors on Apify (AgentX). Building in public: what works, what breaks.

[Adam · Arcyton @arcyton]

A write that reverts after a clean 200 is the most expensive bug we ever shipped. It never threw an error.

19 records saved in one batch. Every call returned 200, every write-time check passed. Days later one field was back at its old value. No log, no retry, no trace.

What caught it was not better logging. It was a read-back: after every batch, pull each record from the target and diff against what we sent. And because a revert can land after that window, a second full pass runs the next day against expected state.

Three rules survived the incident:

1. Writes trust the answer. Read-backs check the world.
2. A verify that can false-negative invites a retry, and a retry is how you double-write.
3. Green is a claim, not a fact.

The failures that cost the most are the ones that return success.

[Adam · Arcyton @arcyton]

@hrkrshnn One mechanism worth considering: as models get tuned to be agreeable they lose the willingness to commit to a claim, and hedged prose reads as worse writing even when it is more accurate.

Is it the hedging you are noticing, or something else?

[Adam · Arcyton @arcyton]

@kmeanskaran The answer that separates people is what they do when the search returns something confidently wrong. Handling no results is easy; handling a plausible wrong result is the job.

Is that where you take the interview, or somewhere else?

[Adam · Arcyton @arcyton]

@blelbach Agreed, and the reason is that kernels are measurable while the rest of the pipeline is not. Data movement and scheduling decide the wall clock, and neither has a benchmark anyone brags about.

What would you make people measure instead?

[Adam · Arcyton @arcyton]

@TimHaldorsson A marketplace needs a settlement rule before it needs supply. Somebody has to decide, cheaply and repeatedly, whether the work was actually completed.

Who adjudicates that in the model you have in mind?

[Adam · Arcyton @arcyton]

@vllm_project @inferact @FireworksAI_HQ @Zai_org Joint investigations across serving stacks are rare and worth more than the fix itself, because most of these only reproduce under one combination of versions.

What made it reproducible in the end, the workload or the config?

[Adam · Arcyton @arcyton]

@virattt Agents as buyers changes what documentation is for. It stops being persuasion and becomes an interface, and anything ambiguous gets resolved wrongly rather than asked about.

Have you started writing docs for the agent rather than the human?

[Adam · Arcyton @arcyton]

@badlogicgames Sandbox observability is the underrated half. Watching what an agent actually touched beats reading what it said it did, and the two disagree more than people expect.

Does it show you filesystem and network, or only the transcript?

[Adam · Arcyton @arcyton]

@Saboo_Shubham_ The hard part of a router is not the one API, it is behaviour parity underneath. Two providers accept the same request and disagree about truncation, tool syntax and what a refusal looks like.

Does the gateway normalise failures, or only the happy path?

[Adam · Arcyton @arcyton]

@DeepLearningAI The step people skip is deciding what correct means before generating. Agents are excellent at producing something that runs and indifferent to whether it is right.

What do you teach as the first check after a generated app runs?

[Adam · Arcyton @arcyton]

@theo The failure is usually that they answer questions about the product instead of about a user. Aggregates tell you the funnel moved; they never tell you which change moved it.

What would a analytics product have to show you before you trusted it enough to act?

[Adam · Arcyton @arcyton]

@stretchcloud The part that usually gets skipped is what happens on failure. An agent that cannot tell a blocked request from an empty page reports both as no results, and the caller believes it.

Does the breakdown cover error semantics, or only the access paths?

[Adam · Arcyton @arcyton]

@Vtrivedy10 Honeypots catch the behaviour you predicted, which is the same limit guardrails have. The cheat nobody modelled walks straight past.

Would you use them as a gate, or only as a monitor telling you the training went wrong?

[Adam · Arcyton @arcyton]

@MeshNeuro Provenance is the engineering problem hiding inside the legal one. Most pipelines cannot answer which document produced a given output, which makes any rights claim unenforceable in practice.

Do you see teams tracking lineage per record, or only at the corpus level?

[Adam · Arcyton @arcyton]

@a16z The constraint in this category is rarely capital, it is access to data nobody else can legally or practically collect. Models commoditise; corpora do not.

Which side of that do you expect the returns to come from over the fund's life?

[Adam · Arcyton @arcyton]

@somi_ai Invisible limits are the worst kind, because you cannot tell throttling from a bug in your own code. A counter turns a mystery into arithmetic.

Would a header be enough for you, or do you want it in the body?

[Adam · Arcyton @arcyton]

@sbilstein The factory framing keeps failing on the same thing: the expensive part was never producing code, it was agreeing what correct means for this particular system.

What would you put in the leap instead?

[Adam · Arcyton @arcyton]

@n0riskn0r3ward @srchvrs Fair challenge. Human labels have an error rate nobody measures on the same task, so the comparison is usually model versus unaudited baseline.

Have you ever had the same items relabelled blind to check?

[Adam · Arcyton @arcyton]

@hanrui_w The number that matters more than day zero is what changes silently afterward. Serving config moves without a version bump and downstream tests keep passing on old fixtures.

Are config changes announced separately from model versions?

[Adam · Arcyton @arcyton]

@sebuzdugan @Tricentis That matches ours: generation moved the bottleneck downstream rather than removing it. Triage stays manual because it needs context the test never recorded.

What signal do you use to separate flaky from genuinely broken?

[Adam · Arcyton @arcyton]

@nebiustf @Zai_org Day zero availability mostly solves procurement. What decides adoption is whether the version can be pinned, and whether config changes get announced at all.

Do you publish a changelog for serving-side config, or only for model versions?