[Requests and ideas] Scheduled workflows should let you select the agent and the context tier (per workflow)

[yoyogin]

Mood: 😔
Category: Requests and ideas

Scheduled workflows should let you select the agent and the context tier (per workflow)

Describe the feature or problem you'd like to solve

Scheduled/recurring workflows (automations that run a saved prompt on a timer) currently offer no control over two settings that decide whether a run can even fit in context:

1. Agent — every scheduled run executes as the built-in default agent, which loads the full set of configured MCP servers. There's no way to point a workflow at a lean custom agent with a scoped tool allowlist.
2. Context tier — every scheduled run uses the default context window (e.g. 200K), even when the selected model supports a much larger tier (e.g. ~1M / long context). The interactive model picker can switch tiers, but that choice does not propagate to scheduled runs.

The result is context exhaustion before useful work starts. With a broad set of MCP servers configured, the tool definitions alone can consume most of the input budget. Two real stuck runs:

• Tool-definition tokens ~186K of a 200K window → only ~1K left for the actual conversation; total tokens then exceeded the limit.
• Tool-definition tokens ~165K; the run climbed to ~198K/200K over a few hours and was killed.

Because a workflow can't select a lean agent (to cut the tool roster) or a larger context tier (to absorb it), these runs get stuck. Interactive sessions work around this by choosing an agent and tier by hand — scheduled runs can't.

Proposed solution

Add two optional per-workflow settings, exposed in the workflow editor and in whatever store backs workflows:

• agent — select any built-in or installed custom agent for the workflow, so its MCP tool surface (and instructions) apply to the run. Defaults to the current built-in agent.
• contextTier — select the context window tier (e.g. default / long_context), matching what the interactive model picker already exposes. Defaults to current behavior.

Either one alone helps materially; together they let an automation run with exactly the tools it needs at a context size that fits.

Example prompts or workflows

• A daily maintenance automation that needs only one or two MCP servers: point it at a lean agent → tool definitions drop from ~165K to a few K, leaving room to actually work.
• A long-running investigation automation: set contextTier: long_context so a multi-hour run doesn't hit the ceiling mid-task.

Additional context

Related to #3525 (context window / reasoning for programmatically-started sessions and subagents) — scheduled workflows are the timer-driven counterpart of the same need. A per-tool / lazy tool-loading mechanism would reduce the pressure further, but per-workflow agent + tier selection is the smaller, targeted fix.

---

Field	Value
App version	1.0.14
OS	Windows 10.0.26200
Theme	GitHub
Path	/chat
Tenure	Week 3