Next.js instrumentation starts embedded AI, workflow, email, and search workers for development. Hosted Ralti uses independent supervised cloud workers with ATLAS_WORKER_MODE=external. Run the agents role for persistent AI teammates; do not assume opening a browser or the development web process starts that role. Web and worker services share their restricted database, private storage configuration, encryption secret, and approved provider configuration.
npm run worker -- --role=all
# Isolate roles when operating separate supervised processes:
npm run worker -- --role=agents
npm run worker -- --role=email,workflows,search
# A controlled single pass against an isolated test environment:
npm run worker -- --role=search --onceSaved work survives disconnected clients#
AI requests are actor/workspace-scoped database jobs. Creation idempotency keys prevent duplicate requests after lost responses. Interrupted jobs retain saved progress for recovery. Cancellation and dismissal are distinct from navigating away. The AI job service bounds concurrency, queue size, and the number of active and retained requests per actor; these are capacity controls, not throughput guarantees.
Visual workflow runs are pinned to immutable published versions. The workflow worker saves progress as steps finish and resumes after approvals, waits, and recoverable interruptions. Draft edits do not rewrite a running version. Completed effects remain recorded when a run is cancelled. Existing helpers retain their own event and schedule behavior: helpers start paused, one overdue interval is dispatched after a restart, and helper-produced changes do not recursively trigger other helpers.
| Control | Effect |
|---|---|
| ATLAS_DISABLE_WORKER=true | Disables embedded workflow and search worker startup |
| ATLAS_DISABLE_AI_WORKER=true | Disables embedded AI worker startup |
| ATLAS_DISABLE_EMAIL_WORKER=true | Disables embedded email and notification worker startup |
| ATLAS_DISABLE_SEARCH_WORKER=true | Disables embedded indexing startup; existing vectors remain searchable |
For external workers, pause the supervised process or omit its role; do not treat embedded startup flags as a global kill switch. Use PostgreSQL to coordinate separate processes. Adding workers without accounting for database pools, provider quotas, and file access can reduce reliability instead of improving it.
Always-on does not mean continuous model calls#
Persistent agent responsibilities and individual execution runs are separate. Assignments, mentions, selected changes, schedules, and future follow-ups wake eligible work. Waiting releases compute. Durable state and leases coordinate recovery across worker restarts; one reasoning run executes per agent at a time. Dispatch and effect commits recheck authority, cancellation, and current versions.
Agent daily workspace and agent allowances reset at 00:00 UTC. An otherwise affordable check saves progress and wakes after the reset; its originating-work budget spans the entire causal chain. Unknown pricing, unreconciled usage, or uncertain external effects block for review. Pause, stop, revoke, and deliberate catch-up have different effects.
Calendar schedules use calendar periods#
Flows support actual daily, weekly, and monthly schedules in an IANA time zone, including the last calendar day. They claim one occurrence per Flow period and may catch up within six hours; older periods are not backfilled. Publication activates the configured trigger. This differs from older helper interval recovery, which dispatches one overdue interval.