How a Play is born
“Machine intelligence is going vertical.”
— Ben Horowitz et al., The Machine Age Fund
Compute can scale output faster than headcount. Without expert context, it can also scale rework.
The metric that matters is judgment per compute cycle.
The expert supplies that judgment: domain knowledge, constraints, corrections, and the definition of done. Rote lets the agent act on it, preserves the successful path, and turns it into a Play the team can inspect, run, and improve.
The chain
01EXPERT → AGENT → ROTE → PLAY02judges acts records carriesA Play begins with your judgment. You define the useful result, correct mistakes, set boundaries, and decide when the work is complete.
The agent turns that direction into actions. It searches, calls tools, checks results, and changes course when an attempt fails.
Rote gives the agent a recorded way to reach APIs, local tools, and websites. It keeps every response and the connection between one action and the next.
The Play is source-visible code compiled from the successful trace. It carries the method, parameters, dependencies, permissions, and presentation into future runs.
| Stage | What it contributes |
|---|---|
| Expert | Intent, correction, constraints, and acceptance |
| Agent | Reasoning, tool choice, recovery, and execution |
| Rote | Recorded actions, addressable evidence, and trace structure |
| Play | An inspectable method that can run again with fresh inputs |
Three ways Rote reaches the world
The agent can use three kinds of recorded action. All three write into the same workspace.
Adapter · APIs
An adapter is configuration that tells Rote how to call an API. Rote can work from OpenAPI, GraphQL, gRPC, Google Discovery, or a live MCP server.
The adapter describes available operations, authentication, request parameters, and response shapes. Each call returns a reference the agent can query and use later.
Proc · CLIs and the local computer
Proc runs a local command or another CLI through Rote. It records standard output, standard error, exit status, duration, and related files.
Use it for tools such as gh, kubectl, package managers, deployment CLIs, and scripts already available on the computer.
Browser · Pages and tabs
Browser opens a browser run and interacts with tabs, pages, forms, and controls. It records page snapshots and extracted results beside API and process responses.
Use it when a service has no suitable API, or when the task depends on a signed-in web interface.
The context workspace
Consider this request:
Fetch recent issues from Datadog. Find the deployed version, identify its GitHub release patches, and add the related Linear triage details.
One request can cross an API, a deployment CLI, GitHub, and a browser run. Rote records every request and response in one workspace as references such as @1, @2, and @3.
Later actions point to those references instead of copying entire responses. The deployed version can feed the GitHub lookup, whose issue identifiers can feed the Linear query.
Those connections form the trace graph. The workspace preserves both the evidence and the path that produced the result.
From a successful trace to a Play
After you confirm the result, the agent reviews the trace with your corrections in view. Failed branches and retries remain available as evidence, but they do not become the default method.
Rote can then crystallize the successful path into a Play. It removes dead ends, turns fixed values into parameters, and preserves the dependencies between steps.
You inspect, lint, and test the generated Play before releasing it. A released Play has a versioned URI that another permitted agent can inspect and run.
On the next run, the agent supplies fresh inputs to the Play instead of rediscovering the same tool path. If a service changes, the failed step shows where to repair and version the method.