Roko PlatformDocs

Implementation

Background agents

Cron agents read the project on a schedule and propose what they find as new change requests.

Purpose

Some work never reaches the top of anyone's list: a missing test for a failure path, a dependency with a known weakness, a decision the code makes that the working model never recorded, a page that slowly got heavier. Teams that wait for someone to notice these usually notice them late. A regular, narrow check finds them while they are still small.

Roko Platform runs such checks as cron agents. Each one reads the project on a schedule, looks for one kind of problem, and proposes each problem it confirms as a new change request. It never changes code or the working model itself. People review its change requests like any other.

The cron agents

Mon · 09:00

Test Coverage Analyst

Missing high-value tests the Quality Plan calls for.

0 9 * * 1

Tue · 09:00

Security Auditor

Security weaknesses in the working model and the code.

0 9 * * 2

Wed · 09:00

Librarian

Context the code makes clear but the working model does not record.

0 9 * * 3

Thu · 09:00

Frontend Performance Analyst

Frontend performance problems, measured on a local build.

0 9 * * 4

Fri, Sat, Sun: no runs by default.

The scheduler checks every minute which agents are due.
An agent run reads the project and its code, for up to one hour.
New change requests arrive In review, or none at all.
The default schedules, all at 09:00 UTC. A project can change each schedule, or turn an agent off, on its Agents page. Each run can only open new change requests.
AgentDefault schedule (UTC)What it looks for
Test Coverage AnalystMonday 09:00Missing high-value tests that the Quality Plan calls for.
Security AuditorTuesday 09:00Security weaknesses in the working model and the code. It does not scan or fuzz. It may run the dependency audit and the project's tests. At most five change requests per run.
LibrarianWednesday 09:00Context that the code makes clear but the working model does not record. It drafts a scoped change request for each gap.
Frontend Performance AnalystThursday 09:00Frontend performance problems. It reads the frontend, then measures the suspect pages against a local build with Lighthouse and Playwright, never against a deployed environment. At most three change requests per run.

Before it proposes anything, each agent lists the project's change requests to skip problems that are already proposed. A run that finds nothing new opens nothing. That is a normal result.

What they produce

A cron agent writes only new change requests, and only change requests it created in the same run. Its instructions allow three tools for that: create_change_request, set_change_request_section, and upsert_change_request_model_change. The change requests arrive In review and appear in the project's Change requests list with the rest.

Scheduling and configuration

Open Agents in the sidebar, under Implement. The Registry tab lists the Cron agents and the Workflow agents, with the columns Agent, Schedule, Model, History, and Last run. Each cron agent row has these controls:

ControlWhat it does
Enabled / Disabled toggleTurns the agent on or off for this project. Agents are on by default.
Edit scheduleOpens the Schedule dialog. Enter a cron expression in UTC, or pick one of the Presets, then press Save schedule.
Edit modelPicks the model this agent uses. On its default, it uses the project's default model.
RunStarts a run now, outside the schedule.
View promptShows the agent's system prompt. The prompt is defined in code and is read-only.

The Runs tab lists every agent run in the project. Each run has its own Agent run page with its transcript.

How a run happens

  • The platform checks every minute which agents are due in each project. It skips archived projects and disabled agents.
  • It does not make up runs it missed, for example while the platform was down.
  • A due run joins the same queue as implementation runs and has the same one-hour time budget. See Budgets.
  • A cron agent needs a model, so the project needs a model provider. See Model provider.

Best practices

  • Fill in the working model first. The agents compare the code against the working model. Test Coverage Analyst reads the Quality Plan, and Librarian looks for what the model does not record. With empty documents, they have little to compare against.
  • Triage their change requests like any other. Accept what is worth doing, and close the rest with a comment that says why. The next run reads the existing change requests and skips duplicates.
  • Turn off what does not apply. A project with no frontend has no use for Frontend Performance Analyst.
  • Spread the schedules. The defaults run one agent per weekday so their change requests do not arrive at once. Keep that spacing if you change them.
  • Read the prompt before you rely on an agent. View prompt shows exactly what the agent checks and what it ignores.