Sierra's researchers ran function-calling agents through customer-service tasks with real rules. The best of them, gpt-4o, succeeded on fewer than half. Asked to do the same task eight times, it got all eight right on fewer than a quarter. An agent that works in the demo is a turkey on day thirty.
[1]Thesis · v3.0
the model is not the control.
An agent that succeeded a thousand times has told you nothing about the thousand-and-first. So the things that must never go wrong are not decided by the model: not what it may do, not what it may spend, not what counts as a fact, and not whether the work is done. Decide those outside it, and you can hand it almost any goal.
Abstract
Abstract
Companies are about to let software spend their money and speak in their name, and then they'll want to hand it whole goals: launch in a new market, finish a recall, keep a directory true. The usual answer to "is it safe" is a better model. This paper argues the question is aimed at the wrong component.
A model's behavior is a distribution, and you can't put a distribution in charge of a bank account, or let it mark its own homework. What an agent may do, what it may spend, which actions wait for the owner, what gets written down, what it takes as fact and what counts as done have to be settled outside the model, in code that runs before the side effect and after the work.
Get those right and you can hand over more than single actions. Anything whose result can be checked can be handed over as one goal. Whatever the budget can't cover goes back to the goal's owner, at that step and nowhere else. Below are seven theses, the published evidence behind each, how OneShot enforces it, and where it stops working. We're unsure about most of the future of agents. We're sure about the part that is enforced.
- action
- A call with a side effect outside the system: an email sent, a call placed, a purchase made. Reading is not an action.
- policy
- A rule evaluated in code before an action runs. It returns allow, deny or hold. The model never sees it and cannot argue with it.
- spend
- Money committed by an action, quoted in USDC before the call and capped per call and per day.
- approval
- A human decision bound to one specific held call, used once.
- receipt
- A signed record of one action that a third party can verify without asking us.
- observation
- A fact about one entity, with the date it was seen and the source it came from, recorded outside the model. History, comparison and watching are all read from these.
- goal
- An objective with a budget and a done-check, run as many actions over days or months. It has a deadline only when time is part of what done means.
- owner
- Whoever delegated a goal and paid its budget: a person, or the agent whose own goal this one serves. The owner can approve and confirm. The agent doing the work can't.
- done-check
- A test of whether a goal succeeded, written before the goal runs and evaluated in code. The model can propose that a goal is done. Only the check, or the owner, can confirm it.
Seven conditions
what the seven make possible
Each thesis is a control. Read together, they are also the conditions for handing an agent a whole goal.
| To hand over a goal, it has to be | Thesis |
|---|---|
| bounded | T1 permissions are not a prompt |
| priced before it runs | T2 skin in the game, in USDC |
| approved by its owner where it goes past the budget, and only there | T3 approvals are for the irreversible |
| on the record | T4 a receipt you don't have to trust us for |
| payable without a person at the till | T5 payment is a protocol, not a wallet we hold |
| checked by something other than the model | T6 done is not the model's call |
| judged on facts the model did not write | T7 memory is not the model's either |
A launch film meets all seven with no approvals: its budget covers every step. A satellite launch meets all seven with a handful: the commitments too large for any budget its owner would sign in advance, like the launch contract. The difference between them is how many commitments outrun the budget, not a difference in kind.
01 · The evidence
a thousand good days prove nothing
A turkey is fed every day for a thousand days. Every feeding raises its confidence that the farmer means well. Day 1,001 is the Wednesday before Thanksgiving.
Taleb borrowed the bird from Russell, who used a chicken. The point survives the change of species: a record of success says nothing about the case that matters.
Text an agent reads can carry instructions, and the agent can follow them. Willison's trifecta is access to private data, exposure to untrusted content and the ability to communicate externally. An outbound agent has all three by job description: it reads the prospect's site, and then it sends the email.
[2][3][4]Gartner's forecast names three causes: escalating costs, unclear business value and inadequate risk controls. Forecasters are paid to sound sure, so read the number as a mood. The third cause is the only one an engineer can do anything about.
[5]02 · Seven theses
what gets decided before the model runs
T1 permissions are not a prompt
What an agent may do is decided by a policy outside the model, on every call, including the calls a goal makes on its own.
Tell a model "never email more than ten people" and you have made a request. A policy that returns 403 is a fact. Most of agent security is the work of turning the first into the second.
OWASP files this under Excessive Agency and says it without decoration: implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed. NIST opened a project on agent identity and authorization in February 2026 for the same reason.[6][7]
How OneShot does it
On every route that reaches the outside world on someone's behalf (email send, SMS send, voice call, commerce buy, physical mail, LinkedIn invite, withdraw, react, reply and message, website build, browser task), the action policy runs once the price is known and before any money moves. It returns allow, deny or hold.
A test fails the build if a paid route with a side effect is neither gated nor on the read-only list.
A denied or held call is never charged. If the policy of an agent known to have one can't be read, the call is held. It fails closed.
Calls a goal makes for itself pass the same policy as calls an agent makes directly.
VerifyA denied call returns 403 action_denied_by_policy and leaves no charge on the ledger. A goal's email under a deny rule is refused before dispatch, and nothing is queued or charged.
T2 skin in the game, in USDC
Every paid action is priced before it runs and capped before it signs.
Hayek's point about prices was that one number carries more information than any planner can. It turns out the math doesn't check your species. An agent that sees the price picks the $0.02 email over the $0.35 call without being told.
Seeing the price is one thing; being stopped by it is another. A budget reconciled at month end tells you what already happened. The cap has to bite before the payment is signed.[8]
How OneShot does it
Every paid call is quoted first: the x402 response carries the price, and the agent signs only that amount.
Each agent can carry a per-transaction cap and a daily cap in USDC. Both are checked against the receipts ledger, plus calls still in flight, before payment verification. An over-budget call never signs.
An agent with a cap stays capped during an outage: if its budget can't be read, the call is held.
A goal is priced per unit from the price list. The model only proposes what a unit needs. Code can use measured attempts to set the expected price; today's quotes still use the attempt cap and mark it unmeasured. Our own failures aren't charged.
VerifyGET /v1/agents/me/budgets returns the caps and today's utilization, ledger and in-flight separately.
T3 approvals are for the irreversible
A person approves the actions that can't be undone, and only those. Inside a goal, the budget approves everything it can pay for, and the owner signs only what goes past it.
Taleb calls harm done by the healer iatrogenics. A person asked to approve a $0.0004 email isn't doing oversight. They are being trained to click yes, and by the hundredth click the approval means nothing.
The EU AI Act asks for something narrower and harder: a person able to override the system, or halt it with a stop button. The Act wants someone with the power to say no. Asking for yes a thousand times a day wears that power down.
A launch takes months of meetings, and most of them are verification done in advance for steps that can't be taken back. Hand that goal to an agent and the reversible work can run on its own. The budget already approved most of what it buys along the way (parts, filings, freight) when the owner paid for the goal. What's left for the owner is the short list of commitments bigger than anything they agreed to in advance: the launch contract, a deposit past the threshold they set. A meeting is an approval that hasn't been compressed yet.[9][6]
How OneShot does it
A policy can hold a call for approval. The approver is paged once per distinct call. The approval is bound to the agent and a hash of that exact request, expires, and is consumed atomically, so it buys one execution and no more.
Until it is approved, the held call costs nothing and does nothing.
A goal can carry its own holds: any single call over a spend threshold, or any tool the owner lists, waits for the owner. The approval covers that tool with those exact parameters at no more than that amount, and it's used once.
A goal that runs out of budget goes back to its owner at that step, with the units left and the expected cost to finish. An open goal gives the price per unit instead. If we can't price the remaining work, we say so. The goal waits for funding; if the hold expires, it ends without claiming to be done.
VerifyWithout an approval the call returns 403 approval_required with the pending approval. Replaying a used approval is refused.
T4 a receipt you don't have to trust us for
Every action leaves a signed record that anyone can check offline.
Pacioli wrote down double entry in 1494. It has outlived every merchant who used it. A ledger is the most Lindy thing in commerce.
Article 12 of the AI Act asks high-risk systems to record events automatically. But a log we keep proves only that we kept a log. A signature proves it to someone who has every reason to doubt us.[10][7]
How OneShot does it
Every terminal receipt is signed with Ed25519. In production an unsigned receipt counts as an incident.
The signer and the verifier share one canonical payload, so they can't drift apart. Verify with the SDK's verifyReceipt() or the oneshot-verify-receipt CLI against our public key. No call to us needed.
A goal's closing receipt includes its done-check and the result, signed with the rest.
Verifynpx oneshot-verify-receipt <receipt.json>
T5 payment is a protocol, not a wallet we hold
Agents pay over open standards, so the rail outlives the vendor.
HTTP has carried a status code for machine payments since 1999. 402, Payment Required, sat reserved for future use for over twenty-five years because nobody had a buyer who wasn't a person clicking a button.
Now there are two serious answers. x402 is an open standard for payment over HTTP with a foundation behind it. Stripe's Agentic Commerce Protocol, built with OpenAI, keeps the merchant of record in control of what is sold and how. Pick the rail most likely to outlive whoever sold it to you.[8][11]
How OneShot does it
OneShot takes both. USDC on Base over x402, verified and settled on-chain. Cards over Stripe ACP checkout sessions, paid with a Stripe shared payment token.
Neither rail requires you to park money in an account we control.
VerifyEvery paid route answers an unpaid request with 402 and a price you can read before you pay.
T6 done is not the model's call
Whether a goal succeeded is decided by a check written before it runs, or by its owner, never by the model that ran it.
An agent asked whether it's finished will usually say yes. It isn't lying. It's doing what it was built to do, which is produce a plausible next step, and "done" is always plausible. A model grading its own work is the turkey again: a thousand confident completions, and nothing in them about the one that wasn't.
The fix is old. A purchase order lists what will be delivered before anything ships, and the goods are checked against it at the dock. Write down what done means before the work starts, and check it with something, or someone, that didn't do the work: a count of meetings on the calendar, delivery statuses from the carrier, a verification result.
This is also how judgment gets handed over. Nobody can check whether a launch film is good. You can check whether every claim in it is on screen and whether it meets a spec someone wrote once. The judgment goes into the spec. After that, the work can be delegated.
Anthropic's advice to agent builders is to start with the simplest solution and add complexity only when needed. A done-check is the simplest form of that: one workflow, and a definition of success agreed before anything is built.[12]
How OneShot does it
A goal carries a done-check from the moment it's quoted: a count over its own tasks, events or emails, or a tool result.
When the planner says a goal is complete, the check runs in code. If it passes, the goal completes. If it fails, the planner gets the missing condition and keeps working within the budget, and after the last attempt the goal fails and says what's missing.
A goal without a check can't complete on the model's word. It waits for its owner to confirm, and the owner can be a person or the agent that delegated it. The agent that did the work can't confirm it.
A deadline, if the goal has one, is part of the check. Once it passes nothing new starts, and the check runs once more to decide.
A goal that runs out of budget waits for its owner, never counts as done. Units the persistence policy could not reach are reported as unreached in the check's result.
Enterprise pilots start from one workflow and the measures that will judge it, written down before we build.
VerifyA goal whose check fails ends failed with the missing condition, not completed. GET /v1/compute/:goalId shows the check, its result and why the goal ended.
T7 memory is not the model's either
What was true, what is normal and what just changed are read from a record kept outside the model, with a date and a source on every fact.
Ask an agent what a supplier charged last March and it will tell you. The number will be plausible, because plausible is the only thing it makes. Asked specific, checkable questions about federal court cases, the models in one study made up the answer between 58% and 88% of the time, depending on the model. A model can reason about last March. It can't have been there.
Give it a search tool and a browser and it can put today together. Slowly, and at a cost, but it can. Three things it can't put together at any price. The past: if nobody wrote down last month's price, it is gone. The whole population: a median needs every peer, and a lookup has one. The watch: noticing a change when nobody asked. The people who built retrieval for these models said in 2020 that providing provenance and updating what a model knows were open research problems. Three years later every model tested still struggled with fast-changing knowledge, whatever its size.
A merchant who keeps no books can still tell you what he is owed. He will be confident, and he will be the only witness. That is where the sixth thesis was exposed: a check that reads facts the worker supplied is the worker marking its own homework by another door. So facts get the treatment actions got. They are written down when they are seen, with where they came from and when, by something that isn't the model.[13][14][15][16]
How OneShot does it
Every answer from a data tool carries its evidence: a source and a date. It is kept, where it can be, as an observation: a dated, sourced fact about one entity. The model reads the record. It doesn't write it.
Changes answers what differs since a date, and says since when it has been recording. For a day before that it has nothing, and says so.
Benchmark places one value among its peers and says whether that population is every peer or only the ones someone has looked up. Across fewer than twenty it refuses.
Watch is the attention. Sources are read on a schedule, and a change is delivered once and charged once, under a spending cap that has to be set when the watch is made. At the cap it pauses and says why.
Who may read a fact is settled in the database. A fact about a person is private to the agent that paid for it unless it comes from a public register.
VerifyA lookup result arrives with evidence: a source and a date. Changes states recorded_since. Benchmark states whether its population is complete, partial or looked_up.
03 · Control map
for the reviewer with a spreadsheet
Our mapping, offered as a starting point for your own. It is not a certification or a compliance claim.
| Thesis | NIST AI RMF | EU AI Act | OWASP LLM Top 10 (2025) |
|---|---|---|---|
| T1 permissions are not a prompt | Govern, Manage | Art. 14 | LLM06 Excessive Agency |
| T2 skin in the game, in USDC | Manage | n/a | LLM06 (excessive autonomy) |
| T3 approvals are for the irreversible | Govern, Manage | Art. 14(4)(d)(e) | LLM06 (human approval) |
| T4 a receipt you don't have to trust us for | Measure, Manage | Art. 12 | n/a |
| T5 payment is a protocol, not a wallet we hold | Map | n/a | n/a |
| T6 done is not the model's call | Measure, Manage | Art. 14 | LLM06 (excessive autonomy) |
| T7 memory is not the model's either | Map, Measure | Art. 14(4)(b) | LLM09 Misinformation |
04 · Where the controls stop
via negativa, applied to ourselves
- The world does not promise to answer
We can't know how many attempts it will take to reach a person. The owner's persistence policy caps those attempts. Units still not reached are reported as unreached, not hidden in a claim that everyone answered.
- Uncapped agents fail open
If the budget store is down, an agent with no cap goes through. An agent with a cap is held, as long as the cache that remembers the cap can still be read.
- Prompt injection is not solved
Policy limits what a hijacked agent can do. It does not stop the agent from being hijacked.
- An approval is only as careful as the person giving it
Holding only the calls that can't be undone keeps each approval rare enough to read. Nothing makes the approver read it. A call held by policy that nobody answers in time is denied, and nothing runs. A goal's hold waits for its owner.
- A receipt proves what happened, not that it was wise
Signatures settle disputes about facts. They don't settle disputes about judgment. A done-check proves the goal met its spec. It can't prove the spec was right.
- A check is only as good as its source
Where the only grader available is another model, the goal waits for its owner instead of completing.
- History starts the day we started recording
Nothing before it exists, for us or for anyone who starts later. Changes says since when it has been recording and has nothing for a day before that.
- A benchmark is as complete as what is recorded
It says whether its population is every peer or only the ones someone has looked up, and it refuses a comparison across fewer than twenty.
- The record is best effort
A lookup that can't be written down still answers. The failed write is logged and the call is not failed, and a tool can be declared as one whose answers aren't kept.
- A watch is paid from credits you hold with us
The fifth thesis says neither rail makes you park money with us. A watch is the exception: nobody makes a request when a change happens, so there is no request to pay on. Each charge still gets a signed receipt.
How sure to be of the evidence
- Gartner's figure is a forecast published in June 2025, not a count.
- τ-bench and AgentDojo measured models available in 2024. Newer models score higher. The gap between pass@1 and pass^k is the finding we rely on, not any one score.
- The legal study asked about court cases, on models available in 2023. We rely on the finding that a model answers what it cannot know, not on either percentage.
- AI Act articles are quoted from an unofficial consolidated text. Obligations for high-risk systems depend on classification and have been subject to postponement. This page is not legal advice.
We can't tell you what agents will be doing in five years. Nobody can, and we'd avoid anyone who says otherwise. We can tell you what ours are not allowed to do today, where their facts come from, who decides when they're finished, and how to check all three.
References
References
- [1]Yao et al., τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (2024) Primary source · checked Oct 4, 2026
- [2]Greshake et al., Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection (2023) Primary source · checked Oct 4, 2026
- [3]Debenedetti et al., AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents (2024) Primary source · checked Oct 4, 2026
- [4]Willison, The lethal trifecta for AI agents (16 Jun 2025) Primary source · checked Oct 4, 2026
- [5]Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 Jun 2025) Primary source · checked Oct 4, 2026
- [6]OWASP GenAI Security Project, LLM06:2025 Excessive Agency Primary source · checked Oct 4, 2026
- [7]NIST NCCoE, Accelerating the Adoption of Software and AI Agent Identity and Authorization, concept paper (5 Feb 2026) Primary source · checked Oct 4, 2026
- [8]x402: an open, neutral standard for internet-native payments Primary source · checked Oct 4, 2026
- [9]Regulation (EU) 2024/1689 (AI Act), Article 14: Human oversight Secondary source · checked Oct 4, 2026
- [10]Regulation (EU) 2024/1689 (AI Act), Article 12: Record-keeping Secondary source · checked Oct 4, 2026
- [11]Stripe, Developing an open standard for agentic commerce (29 Sep 2025) Primary source · checked Oct 4, 2026
- [12]Anthropic, Building effective agents (19 Dec 2024) Primary source · checked Oct 4, 2026
- [13]Dahl et al., Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models (2024) Primary source · checked Oct 13, 2026
- [14]Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020) Primary source · checked Oct 13, 2026
- [15]Vu et al., FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation (2023) Primary source · checked Oct 13, 2026
- [16]OWASP GenAI Security Project, LLM09:2025 Misinformation Primary source · checked Oct 13, 2026

