Skip to content

Verified last-mile execution for agents

Compile a demonstration into a program an agent can invoke. Healthy runs make no model calls. Uncertainty escalates. Humans audit. Computer-use agents are the user of OpenAdapt. They are not the executor inside it.

Call it from an agent See how the compiler works Author a workflow

What the calling agent may do / must not do

May: bind declared parameters, invoke a compiled program, read VERIFIED / HALTED / RECONCILIATION_REQUIRED, supply a missing declared parameter, retry a retryable transport failure, escalate.

Must not: summarize halt as success, resolve an identity or effect contradiction, teach by emitting guessed clicks, or be the sole source of a production demonstration.

Machine contract: agents.txt. Outcome vocab: Run outcomes.

Where OpenAdapt fits

OpenAdapt is the governed last mile an agent calls when the next write has no API. A named human authors the program once. The calling agent operates it. The human returns for identity, effect, and judgment halt, then samples seals.

A strong first workflow has these traits:

  • A person can demonstrate the task from start to finish.
  • The inputs are mostly structured and the business intent stays stable.
  • The application has no practical write API for the required step.
  • A wrong action has an operational, financial, or compliance cost.
  • Another system, account, or session can check the result.
  • The work repeats often enough to justify qualification.

OpenAdapt supports automation teams, BPOs, service providers, and software companies that operate browser, native desktop, RDP, Citrix, or other virtual desktop applications. The daily user is the agent those teams already run.

What the compiler produces

  • An inspectable program

    The compiler turns the demonstrated actions into explicit steps, parameters, guards, and expected state changes. You can review the program before it runs.

    The demonstration compiler

  • Evidence for each consequential step

    A step can carry structural selectors, accessibility data, visual anchors, identity requirements, and a declared result check. The available evidence depends on the application surface.

    The capability ladder

  • A typed outcome and report

    VERIFIED means the declared result check passed. A run that cannot meet its identity, result, policy, or state requirements stops and records the reason.

    Run outcomes

How one workflow runs

flowchart LR
    A([Author]) --> B[[Compile]]
    B --> C[[Qualify]]
    C --> D[[Run]]
    D --> E{Result verified?}
    E -->|yes| F([VERIFIED])
    E -->|no or uncertain| G([HALTED + report])
    D -->|interface changed| H[[Re-resolve from evidence]]
    H -->|resolved| D
    H -->|repair needed| I[[Review repair]]
    I --> C

The healthy path follows the approved program with no generative-model API calls. When the interface changes, OpenAdapt first uses the evidence retained from the demonstration. A configured model can propose a repair when policy permits. The repaired version passes the same review and qualification gates before promotion.

OpenAdapt separates action delivery from result verification. A save message on the acting screen does not prove that the intended record changed. A workflow can verify the result through an API, database, document store, read-only session, or persisted-state reacquisition.

Read about effect verification Read about identity checks

Where it runs

Surface Execution options Evidence available
Browser Local, customer-controlled, or managed for approved workflows DOM, accessibility, visual, geometry, and interaction evidence
Windows, macOS, Linux Local or customer-controlled Native accessibility data plus retained visual evidence
RDP, Citrix, VDI Local or customer-controlled Window-scoped pixels, keyboard, mouse, OCR, and visual anchors

Each workflow is qualified against its exact application, environment, identity rules, result verifier, and deployment boundary. The qualification evidence records the supported scope and its limits.

Sensitive runtime observations stay inside the declared execution boundary. Only an approved, sanitized derivative can cross that boundary.

Compare deployment options Review security and data handling

Inspect the evidence

The public references include the source demonstration, compiled replay, declared verifier, run counts, failure categories, and limits for each claim. The visual demo helps you understand a run. The result verifier and retained report determine its outcome.

Watch the real-application demo Review qualification evidence Read the benchmark methods

Check the current Production status

Choose your next step

  • Call it from an agent

    claude mcp add openadapt, then openadapt quickstart --break-it. Never summarize halt as success.

  • Author a workflow

    A named human captures one browser workflow, compiles it, runs it, and inspects the report. Authority, not daily operator.

  • Prepare a production workflow

    Define the identities, results, failure cases, policy, and deployment boundary for one exact workflow.

  • Review the commercial offer

    See the scope, inputs, evidence, and deliverables for a paid Workflow Qualification Sprint.