Skip to content

AI data market

Agent failure data demand: loud failures need fixes, quiet ones need data

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Failed agent runs create data demand only when they are quiet: the run finishes, no error fires and the business outcome is wrong. Loud failures such as timeouts, missing configuration and sandbox faults need engineering, and harnesses like OSWorld's log them unscored. Quiet failures in business deployments often cannot be trained on under OpenAI's and Anthropic's default terms, so fixing them can require licensed records that reproduce the case.

Agent run

Agent failure data demand

This page covers agent failure data demand: loud failures need fixes, quiet ones need data.Illustrative

Key takeaways

  • Failure volume is a weak demand signal: loud infrastructure failures repeat a few signatures many times and create engineering work, not data purchases.
  • Quiet failures, where a run completes and the business result is wrong, are the ones that can point to missing records, and standard telemetry does not flag them.
  • Track the discovery curve: new quiet failure signatures per thousand runs, after deduplication and a replay on clean infrastructure.
  • By default OpenAI and Anthropic do not train on business customers' data, so failures seen in enterprise use can become specifications for licensed substitutes.
  • Scaffolding fixes many quiet failures without new data, as Anthropic's shop experiment showed, so a demand thesis has to beat that alternative signature by signature.

When does a failed agent run create demand for data?

Only when three things hold: the failure is quiet rather than infrastructural, it recurs across deployments rather than in one customer's setup, and fixing it needs real cases with outcomes that the agent's builder can neither pull from its own trainable logs nor replace with better tools and instructions. Most failed runs miss at least one of the three.

The investor decision is whether growing agent deployment widens the market for licensed business records or mainly widens engineering budgets. Both can grow at once, but they are different businesses, and a company that counts every failed run as a data opportunity is counting the wrong thing.

Testing the claim takes evidence that pitch decks rarely contain: failure logs grouped into signatures, a replay on clean infrastructure, counts of how many separate deployments hit each signature, and the contract status of the traces. What follows covers how to read that evidence, an illustrative discovery curve, and the risks that would cut the link between failures and data spend.

Loud failures announce themselves; quiet ones look like success

Infrastructure failures are loud: they throw errors, stop the run or shift scores across the board, and the tooling to separate them already exists. OpenTelemetry's in-development semantic conventions for generative AI define an exception event for errors that stop an operation from completing, naming API errors, rate limiting, model errors and timeouts as examples.

Benchmark maintainers treat such failures as noise to remove rather than results to learn from. OSWorld's run harness records an attempt that hit an infrastructure error without turning it into a score and marks it retryable, and its README warns that tasks needing a Google account or proxy settings will fail to execute correctly, lowering scores, if those are not configured. Anthropic reported that in internal experiments its most and least resourced setups on Terminal-Bench 2.0 were 6 percentage points apart.

Quiet failures raise nothing. The run completes, the tools report success and the outcome is wrong. The Holistic Agent Leaderboard paper (Kapoor and coauthors, 2025) found that agents with identical accuracy can behave very differently, some taking shortcuts or costly actions such as using an incorrect payment method, and argued that such problems go unnoticed without systematic log analysis.

Anthropic's Project Vend supplies the business version. The Claude-run office shop sold metal cubes for less than it had paid for them, turned down an offer of $100 for a six-pack that cost $15 online, handed out discount codes under pressure from customers and gave buyers a payment account it had invented. None of that would trip an error handler.

Which missing input does each failure point to?

Classify each failure by the task input it lacked: the state the agent needed to see, an action it could not carry out, a constraint it did not know, or knowledge of which outcomes follow which choices. The class decides the fix, and only some fixes involve buying records.

Infrastructure and harness failures sit at the top of the table because they should be stripped out before anyone counts demand. Interface failures are mostly tooling problems. Constraint failures split: a written rule belongs in instructions or a guardrail, while a tacit one has to be learned from cases. Judgment and coverage failures are where real records with outcomes earn a price.

TheAgentCompany benchmark, set inside a simulated software company, shows a version of this split. Its authors found tasks involving social interaction or complex web interfaces especially hard, which points toward interface and tooling work, while data science, administrative and finance tasks scored among the lowest, a gap that may reflect missing domain knowledge as much as interface trouble. The best model it tested, Gemini 2.5 Pro, completed 30.3 percent of tasks autonomously.

The failure-to-input map, with illustrative signatures:

Which missing input does each failure point to?
Failure signature (illustrative)Missing inputCheapest fixExternal data demand?
Timeout, rate limit or sandbox crashNone: the task never ran properlyRetry, capacity, isolationNo
Account or proxy not configuredEnvironment setupFix the harness configurationNo
Grader rejects a correct resultA sound outcome definitionRepair the test or rubricNo; it is evaluation spend
Cannot complete a complex web formAn action the agent can perform reliablyAPI access, better tools, interface practiceRarely
Approves a discount above the limitA constraint, written or tacitWritten rule in instructions or a guardrail; cases if the rule is tacitSometimes
Sells below cost or refuses a good offerOutcome knowledge: which choices pay offCases with decisions and recorded resultsYes
Mishandles a rare exceptionState coverage of that exceptionReal examples of the exception with outcomesYes, when own logs are thin or untrainable

The discovery curve: does more deployment keep finding new quiet failures?

Data demand tracks how fast deployment uncovers new quiet failure signatures, not how many runs fail. A signature is a failure pattern defined tightly enough that two runs sharing it need the same fix, for example the same tool, step, precondition and wrong outcome.

The figure follows an illustrative operator across 100,000 runs. Loud failures account for 6,000 failed runs but only 12 signatures, and all 12 have appeared by run 40,000; after that, loud failures only repeat known patterns. Quiet failures account for just 1,500 failed runs, yet they spread across 180 signatures, and 20 of those first appear in the last 20,000 runs.

Read the two curves differently. A loud curve that keeps climbing means unstable infrastructure, which is an engineering budget. A quiet curve that keeps climbing means the agent keeps meeting business situations it does not handle, which is a candidate for data spend. A quiet curve that flattens means the workflow is narrow and the remaining failures can be fixed one at a time.

In this example loud failures outnumber quiet ones four to one by count, so a dashboard of failure rates would point the budget at the wrong problem. The slope at the right edge, about one new quiet signature per thousand runs here, is the number to ask for.

A line chart of cumulative distinct failure signatures over 100,000 illustrative agent runs: loud failures flatten at 12 signatures by run 40,000 despite 6,000 failed runs, while quiet failures climb to 180 signatures from 1,500 failed runs, adding 20 in the last 20,000 runs.
Discovery curves for loud and quiet agent failures. Illustrative: every value is an example, not a measurement.

Why vendors often cannot train on the failures their customers hit

Because default terms for business use keep customer data out of training. OpenAI says that by default it does not train on inputs or outputs from its business products and API, and uses such data only if a customer explicitly opts in. Anthropic says it does not use data from its commercial products to train its models unless the customer chooses to join its Development Partner Program.

That changes what a quiet failure in enterprise use is worth to whoever builds the agent. A provider, or an agent company whose own contracts follow the same default, may learn from customer reports, evaluations or permitted logs that its agent mishandles freight claims or approval chains, yet the traces that show it may be off-limits as training data. The failure becomes a specification: a precise description of a situation the model has to learn.

A specification can be filled four ways: the customer opts in, the team rebuilds the situation synthetically, the team licenses records from other companies where the same situation occurred with a known outcome, or the team fixes it with scaffolding. Only the third creates external data demand, and it wins only when the other three are blocked, too slow or not faithful enough.

Terms differ by contract and change over time, so treat these defaults as the starting point a diligence question tests, not as a rule for any particular vendor.

Could scaffolding fix quiet failures without anyone's records?

Often, yes, and that is the strongest alternative explanation. In the second phase of Project Vend, Anthropic added a supervising agent that acted as the shop's chief executive; with it in place, discounts fell by about 80 percent and the number of items given away was cut in half. Requiring the shopkeeping agent to check costs and delivery times before quoting also helped. Those gains came from structure around the agent: a supervisor and a required check.

A record collection can also miss what an agent needs. Records show what a person did and sometimes what happened next, but an agent needs the state it would have seen, the rule in force and an outcome logged by a system rather than by the actor. A failure that needs all three can only be addressed by records that keep all three.

So the demand test for each quiet signature is comparative. Name the scaffolding fix, try it on replayed failures, and count the signature as data demand only if the failure survives the fix and the needed cases exist in records with state, constraint and outcome intact.

A failure yield audit an investor can request

Ask the company to run, or show, a failure yield audit with fixed definitions. SourceX has not published results from this design; it is set out here so any agent operator or data supplier can be held to the same terms.

The audit, step by step, with the denominator for each measure:

  • Frame: every failed run in a fixed window, from all deployments, with the window length and deployment count stated.
  • Signatures: failures grouped by tool, step, precondition and outcome, reporting the number of signatures and the rule that defines one.
  • Replay: a random sample of each signature rerun on clean, fully resourced infrastructure; signatures that pass on replay count as loud, whatever the logs said.
  • Loud share: loud failed runs divided by all failed runs, and loud signatures divided by all signatures.
  • Discovery slope: new quiet signatures per 1,000 runs in the last fifth of the window, the figure the curve exists to show.
  • Recurrence: for each quiet signature, the number of separate customer deployments that hit it; a signature seen at one customer is configuration until shown otherwise.
  • Trainability: the share of quiet failure traces the operator may use for training under its contracts, and the share it may not.
  • Fix route: for each quiet signature, the scaffolding fix tried first and whether the failure survived it.

Six developments would weaken the thesis. A company that cannot say which one would hurt it most has not tested its own story.

  • Scaffolding keeps winning. A supervisor and a required cost check sharply cut Project Vend's discounts and giveaways; if that pattern generalizes, quiet failures become engineering work.
  • Customers opt in. OpenAI's terms already allow business customers to share data; if they routinely share failure traces, operators get the cases free.
  • Synthetic reproduction is good enough. If teams can rebuild a failing situation in a simulated environment with a faithful outcome, licensed records lose their edge.
  • Capability gains close whole classes. Each model generation can erase failure signatures before anyone buys data for them.
  • The quiet curve flattens. Agents deployed in narrow workflows may run out of new situations quickly, capping demand.
  • Licensed substitutes hit their own limits. Records that reproduce a failure still need rights, de-identification and a buyer who can check their outcomes.

Diligence questions on failure-driven demand, SourceX included

Ask every company that names agent failures as its demand engine to show the audit above, with definitions it cannot change after seeing the results.

SourceX sits on the licensed-substitute route. Companies that choose to license operating records come to it, and it works from a buyer's written requirement rather than a catalog, looking for U.S. businesses whose records fit what the buyer has specified about scope, language, format and volume, timing and license terms. A precise failure specification is the kind of input that process can act on, though whether a matching business and usable records exist is settled case by case.

None of SourceX's request, deal or commercial figures are public, and nothing here implies them; the company shares company and commercial metrics with investors directly. Its raise is open only to accredited investors under Rule 506(c). Investors can ask for materials through sourcex.si/investors and include a question, for instance how many buyer requests have described a specific agent failure.

Questions for any company in this market:

  • What share of your customers' failed runs are loud, measured by replay rather than by log labels?
  • What is your discovery slope for quiet failure signatures, over what window and how many deployments?
  • How many separate deployments hit your top quiet signatures?
  • What share of failure traces may you legally train on, and what fills the rest?
  • For each signature you count as data demand, which scaffolding fix did you try first, and what happened?
  • How many buyers have described a specific failure in a written data request, counted by independent budget owner?

Methods note: sources and limits of the failure map and discovery curve

The failure-to-input map, the discovery curve and the audit design are editorial tools drafted on 2026-10-27. They draw on primary sources checked that day: OpenTelemetry's generative AI exception conventions and OSWorld's harness code and README, read in their GitHub repositories; Anthropic's infrastructure-noise post and two Project Vend reports; the Holistic Agent Leaderboard and TheAgentCompany papers on arXiv; OpenAI's enterprise privacy page; and Anthropic's privacy center article on commercial data use.

Every number in the discovery curve and every failure signature in the map is invented to show the mechanism. Not measured: the loud share, discovery slope or recurrence for any real operator, the share of enterprise traces that cannot be used for training, and any SourceX request or deal figure. Sources verified only through search excerpts should be opened by a reviewer before publication.

Frequently asked questions

Do failed AI agent runs make good training data?

Some do. Failures caused by timeouts, missing configuration or broken graders teach nothing about the task and should be removed first. Failures where the run completed and the business outcome was wrong can be valuable, provided the trace keeps the starting state, the rule in force and the recorded outcome, and the operator may legally train on it.

What is infrastructure noise in agent evaluation?

It is score variation caused by the environment rather than the agent: resource limits, network settings, flaky services or harness bugs. Anthropic reported a 6-point gap on Terminal-Bench 2.0 between its most and least resourced setups in internal experiments. Harnesses such as OSWorld's now record infrastructure errors as unscored, retryable attempts.

Can AI providers train on failures from business customers' agents?

Under their published defaults, OpenAI and Anthropic do not train on business customers' data unless the customer opts in, through data sharing at OpenAI or a development partner program at Anthropic. Individual contracts can differ, so check the specific terms. Where training is off the table, a failure becomes a specification for some other fix.

What is a failure signature?

A failure signature is a pattern defined tightly enough that every run sharing it needs the same fix, typically the same tool, step, precondition and wrong outcome. Counting signatures instead of failed runs stops one repeated infrastructure fault from looking like thousands of separate problems.

Does SourceX publish how many data requests describe agent failures?

No. SourceX's request, deal and commercial figures are not public. The company shares company and commercial metrics with investors directly, and materials can be requested at sourcex.si/investors. Any offering is made only to verified accredited investors.

Sources

  • OpenTelemetry's GenAI conventions (status: development) define an exception event for errors such as API errors, rate limiting, model errors and timeouts that prevent an operation from completing. Source
  • OSWorld's harness records failed attempts without turning infrastructure errors into scores; unscored attempts remain retryable. Source
  • OSWorld's README warns that tasks needing Google account or proxy configuration fail to execute correctly, lowering scores, when not configured. Source
  • Anthropic found a 6 percentage point gap between its most and least resourced Terminal-Bench 2.0 setups in internal experiments. Source
  • The Holistic Agent Leaderboard paper reports that agents with identical accuracy can behave very differently, including shortcuts and costly actions such as incorrect payment methods, and calls for systematic log analysis. Source
  • In Project Vend's first phase, Claude sold metal cubes below cost, declined $100 for a six-pack costing $15 online, gave out discounts and gave customers an invented payment account. Source
  • In Project Vend's second phase, a supervising CEO agent cut discounts by about 80 percent and halved items given away; requiring cost and delivery checks before quoting also helped. Source
  • TheAgentCompany found social-interaction and complex web interface tasks especially hard, data science, administrative and finance tasks among the lowest, and the best model (Gemini 2.5 Pro) completing 30.3 percent of tasks. Source
  • OpenAI does not train on business data from its business products and API by default; data is used for training only if a customer explicitly opts in. Source
  • Anthropic does not use data from its commercial products to train its models unless the customer joins its Development Partner Program. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies