The verified fast path for AI agents

Your agent's repeat work, answered in 2 ms.

Wont learns the tool calls your chat, voice and back-office agents make again and again, proves each one against your model, and then answers it instantly. Anything new, and every write, still goes to your model.

Works withOpenAIAnthropicPipecatDocker

Proven repeat: answered by Wont in 1.8 ms New request: goes to your model Any write: always your model

The cost of a repeat

Every repeated lookup is a full model call.

To find an order, a ticket or a booking, an agent asks its model which tool to call, then waits. In our runs that decision took a median of 0.68 to 0.95 seconds, and the same decisions came up again and again.

0.68–0.95 sfor the model to decide one tool call, at the median
8–20%of an agent's model time spent re-deciding the same lookups, on five of six desks
12–26%of its tokens, on the same five desks

Chat and text

Slower answers, larger bills

Every "where is my order" pays for a model round trip that decides exactly what it decided last time.

Voice

Silence the caller hears

On a call, each decision is dead air: about 0.95 s at the median and 2.3 s at the 90th percentile in our live support run.

Back office

Identify, then look up

Tickets, claims and parcels follow the same steps every time: find the person, fetch the record. The model re-decides it on every request.

How it works

Learned from your traffic. Proven before it serves.

Wont is a proxy you run yourself, between your agent and its model. It never guesses: a step is served only after it has matched your model on your own traffic and a person has said yes.

  1. 01

    Watch

    Change one base URL. Every request still goes to your model while Wont records each decision on your machine and counts the ones your agent makes the same way every time, down to where each argument comes from.

  2. 02

    Prove

    On live traffic, Wont predicts each repeated step beside your model and keeps score. A step earns a certificate only when a 95% bound puts its error under 2%, over at least 24 hours, 2 days and 10 users.

  3. 03

    Promote

    Certified is not enough. A person on your team decides to serve a step, with one command, and can undo it the same way.

  4. 04

    Serve

    Wont answers that step in about 2 ms, in your provider's exact format, and your code runs the tool as it always does. A new prompt, model or tool list stops serving until the step is proven again.

Not a cache, and not a router. Wont learns how each argument is derived, so a new customer's ticket number is read from what they wrote, never replayed from the last conversation.

For chat and text agents

Faster answers on the lookups your agent repeats.

A support agent spends much of its day on the same moves: find the customer, fetch the order, check the ticket. Wont takes over the moves it has proven, so your model's time goes to the reply, the judgment and anything new.

  • Exact, not just the tool. Up to 49.3% of an agent's tool decisions predicted exactly, tool and every argument, on conversations Wont had never seen.
  • Values are read, never replayed. Ticket, claim, policy, tracking and booking numbers, dates and phone numbers: 18 formats, each read exactly for 20 of 20 new customers.
  • The reply stays your model's. Wont serves the tool call; your model still writes every word your customer reads.

Model time and tokens Wont would remove, by desk

model timetokens

Exploratory: each desk's share of the agent's whole model bill, replies included, if every proven step were promoted.

For voice agents

Less dead air on every tool turn.

On a call, your agent can't think out loud. Each time it decides which tool to call, the caller hears silence: about 0.95 s at the median and 2.3 s at the 90th percentile in our live support run.

  • Proven lookups answered in 1.8 ms. "Find the caller by the email they just said" is decided by Wont instead of a model round trip.
  • Reads start before the model decides. Wont's pre-fetch cut the email lookup's wait from 235 ms to 0.9 ms on spoken calls.
  • Proven on real speech. 535 of 537 decisions agreed with the model, from 275 callers, with speech-to-text in the loop.

Where Wont sits on a call

callerspeech-to-textWontmodelvoice reply

Measured on spoken calls · Pipecat, Deepgram, Qwen3-30B

Synthetic callers. The agent speaks a short preamble first, so the first sound the caller hears did not change; the silence after it did. Serving the step itself on calls is measured once its certificate passes.

Benchmarks

Exact where it acts. Instant when it does.

Six support desks, a real model, simulated customers, and every plan frozen before the run. The misses are published with the wins.

1.8 msto serve a proven decision
0wrong answers served and 0 writes served, in every run
49.3%of tool decisions predicted exactly on the best desk

Wont against OpenAI's Decisions API

The same 748 retail tool decisions. The Decisions API guesses the next tool on every turn; Wont answers the whole call, tool and arguments, only where it has proof.

60%70%80%90%100%0%25%50%75%100%share of the agent's tool decisions answeredright, where it answersDecisions API on 100% of decisions: right tool 69.8% of the timeDecisions API on 90% of decisions: right tool 74.5% of the timeDecisions API on 75% of decisions: right tool 81.8% of the timeDecisions API on 66% of decisions: right tool 86.9% of the timeDecisions API on 53% of decisions: right tool 93.0% of the timeDecisions API on 44% of decisions: right tool 96.1% of the timeWont on 24.6% of decisions: exactly right, tool and arguments, 98.9% of the timeWont · 98.9%, whole callDecisions API · right tool only

Swipe to see the whole chart.

MeasureDecisions APIWont
Predictsthe next tooltool and every argument
Right, where it answers69.8% to 96.1%98.9%
Time per decision155 ms1.8 ms
Cost per 1,000$0.33$0
Still needs a model callyes, for the argumentsno
Your data leavesyesno

Wont's 98.9% counts the whole call; the Decisions API's scores count the tool only, and rise as its confidence threshold narrows what it answers. Its time was measured over the network from one laptop.

Six support desks

Tool decisions predicted exactly, tool and every argument, on conversations Wont had never seen.

The numbers by desk
DeskExactWrongWhat Wont took over
Airline49.3%2 of 183Booking by code, flight status, flight search
IT helpdesk47.3%0 of 106Employee by email, ticket, their laptop
Insurance37.6%0 of 68Policy by number or email, its claims
Parcel delivery33.7%0 of 60Tracking by number, by order
Retail24.3%0 of 182*Customer by email, order by number
Clinic13.9%1 of 36Patient by ID; refused where the agent varied

* 2 wrong when run live; today's code refuses both. No wrong prediction was ever served: every one was caught in shadow.

Time to a tool decision

Log scale. A served step skips the model call entirely.

How we test

  • A real model (Qwen3-30B) with simulated customers, 200 to 400 conversations per desk, each plan pre-registered.
  • Retail and airline are τ²-bench's environments; the other four we built on its framework. Voice used synthetic callers through Pipecat and Deepgram.
  • The miss we published. Airline started at 4.3%: Wont could not read booking codes or spoken dates. Two readers later, a fresh run gave 49.3%.
  • Without intent labels, Wont grouping requests itself reached 27.5% on retail. Exploratory.

What if Wont is wrong?

It earns every step, and never touches a write.

A wrong answer costs more than a slow one. So Wont serves nothing until a step is proven on your own traffic and a person says yes, and it steps aside the moment anything changes.

Proven per step

A certificate per route: its evidence, the agent version it holds for, and how each argument is derived.

A person promotes

Nothing is served until someone on your team runs promote, and demote undoes it.

Writes go to your model

Cancelling, booking, refunding: always your model's decision. Wont never serves one.

Stops on change

A new prompt, model or tool list, or a sampled audit that disagrees, sends the step back to shadow.

One kill switch

wont proxy kill on stops all serving at once; every request goes to your model.

Your data stays yours

Traffic stays on your machine. Reports and certificates hold no customer values. Your API key goes only to your provider.

0 writes served. 0 wrong answers served. Every disagreement in our runs was caught in shadow, before anything was promoted.

For developers

See it work in two minutes, no keys needed.

The demo runs the whole lifecycle against a fake model, in memory, for eight kinds of support desk. Then put the proxy in front of your own agent, in shadow.

  1. Run the demo. Watch a step get learned, proven, promoted, served and stopped.
  2. Mark which tools only read. Anything unmarked is never served.
  3. Point your agent at the proxy. One base URL, for OpenAI, Anthropic or Pipecat.
  4. Read the report. See what repeats, and watch each certificate fill.

Pre-release: version 0.2.0, with the proxy, installs from the repository on request until it is on PyPI.

# the whole lifecycle in seconds: fake model, in memory
pip install ".[proxy]"
wont proxy demo --scenario all

E-commerce support  [ok]
   customer:  "Hi, where is my order A-5001?"
   learned:   get_order(order_id <- extract:user:id[-1])
   evidence:  300 comparisons, 0 disagreements, bound 1.49%, certified
   served:    get_order({"order_id":"A-5001"})  (0 model calls)
   never served: cancel_order (a write)
Telecom support  [ok]
   customer:  "Hi, my data stopped working. My phone number is (713) 554-6001."
   learned:   find_account_by_phone(phone <- digits(extract:user:phone[-1]))
   ...

Runs as one Python process or the Docker image in deploy/proxy/, published on 127.0.0.1 only. Pipecat has a full voice example in the repository.

Questions

Straight answers.

Every number on this page comes from a run recorded in our repository, with its plan, its data hashes and its misses. Exploratory results are marked.

Does Wont replace my model?

No. Your model answers everything new, everything uncertain, every reply to your user and every write. Wont takes only the steps it has proven and you have promoted.

Is it for chat or for voice?

Both, and back-office agents too: anything that calls an OpenAI- or Anthropic-compatible model to decide its tools. On a call, the time saved is silence the caller no longer hears.

What if it gets a step wrong?

Before promotion nothing is served, so a miss only raises the bar for certification. After promotion a share of requests is still checked against your model, and one disagreement sends the step back to shadow.

What does it store, and where?

Your traffic stays on your machine, in a local store you can turn off. Reports and certificates hold no customer values, and your API key goes only to your provider.

We don't label requests by intent.

Wont works best with a label per request, such as a flow node or a menu choice. Without one, Wont grouping requests itself reached 27.5% on retail in an exploratory test. Inferring the label is next.

Has it run on real customer traffic?

Not yet. Every result here is a real model with simulated customers. A free shadow-mode pilot measures what Wont would do on your traffic, without serving anything.

Run Wont in shadow on your traffic, chat or voice.

A free design-partner pilot. It records, learns and reports, and serves nothing until you promote a certified step.

or write to prem@usewont.com