Guide · Recurring tasks · Model usage
Do I need an AI model to solve the same browser task on every run?
Short answer
A stable task with explicit actions and checks can run as a saved procedure. Ritoko’s direct replay engine does not call an LLM. Recording, repair, host orchestration and external tools that use models can still consume model usage or incur charges.
Updated
Which decisions need reasoning, and which can be saved?
Consider a weekly supplier import. An agent first discovers the form, determines which input field maps to each column and finds proof that a supplier was created. Once those decisions are correct, the next file may require the same actions with different values.
Save the workflow’s parameters, selectors, business key, commit boundary and verification checks. Direct replay then executes that procedure rather than asking the model to discover each step again. A task that requires interpreting new documents, choosing among changing business rules or navigating unfamiliar sites may still need reasoning for each item.
Which parts of a Ritoko run can still cost money?
- Authoring and recording: the client agent uses its subscription or API configuration to solve and save the task.
- Direct replay: the engine executes saved browser, HTTP and MCP steps without its own LLM call. The surrounding agent can still use a model to launch or discuss the run.
- External services: an MCP tool, OCR API or generation endpoint may use a model and charge for every call.
- Repair and host mode: the agent participates again. Host orchestration is not a promise of zero model usage.
- Operation and maintenance: browser runtime, service access, destination limits and a person’s review time still matter.
Ritoko includes no local OCR engine or invoice parser. Its optional document_image returns a downloaded image for the client model to read; each new image still requires interpretation. See the runtime and document-reading FAQ.
How should I measure the saving on my task?
Record authoring time and model usage separately from the repeated runs. Then measure a representative small batch: elapsed time, confirmed items, review items, external service charges and repair effort. Compare complete business outcomes, including manual follow-up.
A useful budget is initial authoring cost plus repeated execution cost plus expected maintenance and review. No percentage saving follows from the architecture alone. A changing page or model-backed extraction can dominate the budget even when the replay engine itself makes no model calls.
When is saving a procedure worth the effort?
Replay fits tasks that recur, accept structured inputs and expose a result you can check. Keep an agent involved when the task is exploratory or its next action depends on judgment. Start with one verified operation and a short batch, then reuse it once the business rules are clear.
Ritoko is not the only way to avoid inference during execution: generated Playwright code and other tools’ documented replay or cache paths also do that. The relevant difference is how you want to maintain the procedure and track its business outcomes.
Sources
- Playwright: generating executable testsRetrieved
- Stagehand: caching actionsRetrieved