
MA Duplex / The adaptive runtime for AI agents
Rent the model. Own the intelligence.
MA Duplex sits in front of any frontier model, learns from every answer it gets and serves what it knows on its own. Your agents get more accurate, more governed and cheaper with every turn.
16 pts
Over fine-tuning
89%
Fewer model calls
0
Training runs
01 / Why a runtime
Why agents need a runtime.
An agent that calls a model for every decision pays for every decision, forever, and learns nothing your company keeps. An agent in operations also has to be bounded: every turn authorized, every action on record. MA Duplex gives agents both.

02 / Day one
What changes on day one.

- 01
Point your agents at it.
MA Duplex speaks the OpenAI and Anthropic APIs your code already uses. Change the address, not the code.
- 02
It asks, then it learns.
The first time it meets a task it asks the model and learns the answer.
- 03
Every turn governed.
From the first request, every turn is authorized and every action goes on record.
- 04
Nothing to train.
No dataset, no GPU hours, no checkpoint to ship.
03 / Cost inversion
The more it works, the less it costs.
The first time it meets a task it asks the model and learns the answer. From then on it serves that decision itself, in milliseconds.
On repeated work it cut model calls by 89% while getting more answers right than a fine-tuned model. Of the 303,176 answers it gave on its own, 99.9% were correct.

11,207
Model calls in 100,000 requests
99.9%
Of its own answers correct
10 of 10
Runs ahead
04 / Learning without training
Accuracy above fine-tuning. Zero training runs.
Fine-tuning means datasets, GPU hours, checkpoints and a model that starts aging the day it ships. MA Duplex learns from live work, one example at a time. On a public benchmark it reached 91.9% with no training run, against 85.8% for a fine-tuned model and 71.3% for the model alone. On repeated work: 98.6%.

91.9%
On a public benchmark, with no training run
85.8%
Fine-tuned model
98.6%
On repeated work
05 / Always current
Change a fact once. Every answer follows.
Across 20,160 answers through every kind of update, it gave zero stale answers. Standard retrieval gave stale ones 57% of the time.
0%
Stale answers, against 57% for standard retrieval
0 of 8,640
Changed facts answered the old way
100%
Of deleted facts kept out
06 / Governed turns
Every turn authorized. Every rule enforced.
Each turn passes one gate before anything reaches a model or a tool. Rules hold on what a request means, not only on how it is worded, and every action is written to a tamper-evident record your auditors can verify.


07 / Reversible
Unlearn anything, exactly.
Any lesson can be removed on request, and MA Duplex returns precisely to the state it was in before it learned it. Nothing else is disturbed. Every change is on record.
08 / Your stack
Works with the stack you already run.

- 01
Any frontier model.
Swap it tomorrow without losing what was learned.
- 02
Drop-in endpoints.
OpenAI and Anthropic compatible, and usable from LangChain.
- 03
SDKs, a command line and an HTTP API.
Python and TypeScript.
- 04
Your memory, graph and tools.
It works with the knowledge and tools your agents already use.
- 05
Your cloud, your premises or managed for you.
In your own cloud account, on your premises, or in a separate environment operated by Machine Autonomy.
- 06
Offline development.
It runs end to end offline, with no API keys. Many agents and multi-step workflows pass through the same gate.
09 / For operators
What operators get.


- 01
A record auditors can verify.
Every turn and every action, tamper-evident and retrievable by API.
- 02
Rules that hold.
Each turn passes one gate before anything reaches a model or a tool.
- 03
Facts that stay current.
Change a fact once and every answer follows. Delete one and it stays deleted.
- 04
Keys, scopes and quotas.
For every team, with usage, in one console.
10 / Where it runs
Its own environment. Or yours.
Each customer gets a separate MA Duplex environment, operated and managed by Machine Autonomy. Or run it in your own cloud account, on your premises or fully air-gapped. Your data and what it learns stay where you decide.
- A separate environment for each customer, managed for you
- Your cloud account
- Your premises
- Air-gapped

11 / What it costs to run
The bill falls as the volume grows.
Every answer MA Duplex has learned is served without a model call, in milliseconds. The savings grow with volume from day one, and nothing is spent on training.
89%
Fewer model calls on repeated work
59%
Of requests on a public benchmark answered with no model call
0
Training runs, datasets or GPU hours
12 / How a pilot runs
How a pilot runs.


01
Your own traffic
MA Duplex goes in front of the model you use today, on real requests.
02
Your rules
Policies and limits set before the first turn.
03
Measure it
Accuracy, model calls and every turn, on record.
04
Roll it out
For your agents, in our environment, your cloud or your premises.
13 / Where it pays off
Where it pays off.
- 01
Customer support
Accuracy that grows every day and a model bill that shrinks.
- 02
Operations
Tickets, orders and requests at volume, every decision governed and on record.
- 03
Finance and compliance
Decisions that follow the latest rule the moment it changes.
- 04
Agents and copilots
Tool choices and routing served in milliseconds.
14 / Proof
Ahead of fine-tuning, without fine-tuning.
Repeated work
98.6% accurate, 16 points above a fine-tuned model.
- MA Duplex98.6%
- Fine-tuned model82.6%
- Model alone66.8%
Nine requests in ten repeating, across 2.4 million decisions. 89% fewer model calls.
Public benchmark
Beats fine-tuning on a public benchmark, with no training run.
- MA Duplex91.9%
- Fine-tuned model85.8%
- Model alone71.3%
59% of requests answered with no model call, 99.6% of them correctly.
Knowledge updates
Lower is better
Zero stale answers through every kind of update.
- MA Duplex0%
- Standard RAG57%
Stale answers across 20,160 answers in 360 knowledge families.
Output rules
Fine-tuned accuracy on strict output rules, without fine-tuning.
- MA Duplex98.1%
- Fine-tuned model97.7%
- Model alone96.2%
Answers that met every output rule. Zero training runs.
15 / Questions
Questions.
Does MA Duplex replace my model?
No. It sits in front of it. Your model handles what is new, MA Duplex handles what it has learned.
What changes in our code?
Very little. Point your agents at MA Duplex instead of the model. It speaks the OpenAI and Anthropic APIs.
Do I need training data?
No. It learns from live work, one example at a time, and from the corrections your team makes.
What happens the first time it sees a task?
It asks the model, returns the answer and learns it. The next time, it can answer on its own.
How quickly does it pay off?
From the first repeat.
Can we change models later?
Yes. Swap the model without losing what was learned.
Where does it run?
In its own environment, operated and managed by Machine Autonomy, or in your cloud account, on your premises or air-gapped. Your data and what it learns stay where you decide.
Can developers work offline?
Yes. It runs end to end offline for development, with no API keys.
Can it forget something?
Yes, exactly, with the change on record.
How do we get it?
Machine Autonomy offers MA Duplex to enterprises and to teams building agents.
Partners


Own the intelligence your agents create.
Run MA Duplex on your own traffic and watch accuracy rise while the model bill falls.