Dashboard agents and custom serverless code
When to use dashboard agents, when to bring a custom runtime, and how their charges differ.
Reviewed October 9, 2026: dashboard agents and custom-code agents have different runtime and billing paths.
Most clients start in Agents → Create agent, add their own instructions, knowledge, and tools, and try a task. Automations run recurring work without a repository. Completed model steps use AI pricing; this path does not add a separate custom serverless runtime charge.
When custom code makes sense
Use Agents → Custom code for a repository and runtime you own. A compatible serverless deployment scales execution to zero between requests. Cold-start time varies with code and dependencies; there is no fixed one-to-two-second guarantee. Stored data, secrets, and optional reserved capacity can still incur charges while no request is running.
Custom code bills request count and execution time, with usage collected in the background and settled through monthly billing. Builds and model calls are separate. Optional always-warm reserves capacity and adds a time-based charge; it does not remove every source of latency. See Pricing.
State is an application responsibility
The platform provides a scoped managed PostgreSQL binding when required by the custom runtime. Your code must decide what to persist, authenticate callers, restrict tools, and resume saved work. A database connection alone does not make arbitrary code durable. SetCloud no longer supplies agent templates.
Dashboard-agent runs start with fresh context. Website chat has separate visitor sessions. To compare earlier results in an automation, include the earlier data in its input or an explicitly configured tool.
Scheduling and recovery
Dashboard automations use a frequency and time zone. Custom repository schedules use the runtime’s published schedule contract and require deployment after a change. Connected code-review tasks have their own read-only review history; they do not push fixes or run repository tests.
Spend controls can pause AI and autonomous work. Background reconciliation takes time, and a paused runtime does not delete retained data or stop every resource cost. Read spend limits, the Agents quickstart, and bring your own runtime.