Workflows

From a request
to a controlled result.

Conversation, projects, files, agent tasks, structured data, review, and feedback stay connected inside one general-purpose workflow.

Follow the Workflow →Open Interactive Tour
Workflow 01 · Surfaces

What you touch.
Not what routes it.

Routing internals are on the Platform page. This is the operator side: the surfaces a request actually moves through, and what each one is for.

01
Enter
Chat or Project
Work starts in a chat session, or inside a project that carries its own files, context, and history across sessions.
sessions · projects · persistent context
→
02
Attach
Files + Data Grid
Uploaded or connected structured data can be inspected, profiled, queried, charted, and exported inside the workspace.
upload · schema · query · export
→
03
Delegate
Agent Subtasks
Work that decomposes can be handed to dynamic agents running their own subtasks, each visible as it progresses rather than as a single opaque call.
dynamic agents · parallel subtasks
→
04
Control
Action Previews
Operations with real-world effect surface a preview the operator must approve before anything executes — not a silent action taken on their behalf.
preview · explicit approval
→
05
Inspect
Traces
Every dispatch — a routed reply, an agent subtask, a fallback — is recorded as a trace the operator can open, not just a result with no history behind it.
dispatch · latency · outcome · errors
→
06
Close the loop
Ratings
A thumbs up, thumbs down, or written note becomes routing evidence — the mechanism behind it lives on the Platform page's Karma system.
explicit feedback · routing evidence
Workflow 02 · Connected Surfaces

What stays connected
while the work moves.

Chat, projects, files, agent tasks, Data Grid operations, traces, approvals, and ratings share one operating context instead of becoming separate manual handoffs.

Working across tabs
  • You pick the model before you know what the job needs
  • A rate limit or an outage stops you until you move it yourself
  • Context is scattered across chats, tools, and folders
  • You copy subtasks between models by hand
  • A bad answer stays a bad answer — nothing learns from it
  • Risky actions happen with no consistent check in front of them
Working in Agent Forge
  • The job is read first, then matched to a model that suits it
  • If one model stalls, a healthy one picks the work up
  • Sessions, projects, and files stay attached to the work
  • Big jobs split across agents and come back together
  • Your rating and your corrections change the next decision
  • A review pass, and your approval, before anything touches a file