Manual start
runsSomeone presses Run and optionally types a note that the flow can read.
trigger.manual
FlowsDraw how the work
Draw how the work
actually moves.
A flow is your procedure, drawn: what starts it, which capabilities run, where a person decides, and what gets posted where. Your manager builds it on a canvas, publishes a version, and the worker executes that version — with every step recorded.
In the portal today
- Trigger
Application created
Start with an application event
- Decision
Human approval
A named reviewer makes the call
- Action
Post to Slack
Share the approved update
Every step is recorded. A human decision is part of the procedure.
CapitalSource Agent Harness
Saved progress. Explicit approvals.
The execution layer that keeps agent work moving—with scoped permissions, spending limits, and a record of the task. Starting with the Processor application-readiness pilot in Flows.
Preview before deliveryStart in shadow mode to inspect readiness evidence and proposed operations follow-ups without sending a message.
Approve the exact actionEach Slack delivery requires approval of the message and destination. Permissions and supporting evidence are checked again before dispatch.
Follow the same taskSee progress and model spending as application information changes. Reassess or cancel a task from its run detail.
Workspace setup is required. Broader fleet support is planned; the pilot assesses readiness and prepares operations follow-ups. Financing decisions remain with people.
What is built
Authored by the manager, not by us
Flows is one of the few things on this site that is finished rather than designed. It is in the portal, it is covered by tests, and the list below is what it does.
Working
The builder and the runner
- A canvas with a typed palette — drag a node in, wire it up, configure it in a side panel
- Drafts autosave as you work; publishing freezes a numbered version you can roll back to
- Validation runs against the graph, so a flow that cannot execute will not publish
- Runs are recorded step by step: what each node received, returned, and how long it took
- An approvals inbox where a named reviewer sees exactly the data the flow is asking about
- Templates to start from, rather than a blank canvas
Not yet
What a flow cannot do
- Event triggers other than application events and manual start — uploads, Slack matches, schedules, and inbound webhooks are in the palette and refuse to run
- Outbound sends on anything but Slack — Telegram, WhatsApp, and Google Chat connect but cannot be posted to from a flow yet
- Handing a step to a deployed agent, which waits on hosted deployment existing at all
- Five of the seven capability nodes return a contract-locked pending envelope instead of real analysis
The palette
Every node says whether it runs
A node marked coming soon is visible in the builder and refuses to execute, so a flow that depends on one cannot publish. These labels come from the same catalog the palette reads. End to end today that means: a trigger, logic, a human approval, and a Slack message — the other connectors are defined and not connectable.
Triggers
1 of 5 run todayDocuments uploaded
coming soonFires when files land from the portal, a secure link, or a website form.
trigger.documents_uploaded
Slack message matched
coming soonFires on a message in a connected workspace that matches a pattern.
trigger.slack.message_matched
Schedule
coming soonRuns the flow on a recurring clock rather than an event.
trigger.schedule
Inbound webhook
coming soonAnother system posts a payload and the flow starts from it.
trigger.webhook
Capabilities
2 of 2 run todayScore CSG-1003 completeness
runsPercent complete, missing field ids, and gating status for an application.
capability.score_csg_1003_completeness
Generate credit memo
runsA risk-scored memo in markdown, with structured signals and a PDF.
capability.generate_credit_memo
Logic
3 of 3 run todayIf / else
runsTwo branches from one condition over anything upstream produced.
logic.if
Switch
runsMany branches keyed off a value, with a default arm.
logic.switch
Wait
runsPause the run and pick it up later without losing state.
logic.wait
People
1 of 1 run todayHuman approval
runsStops the run, shows a named reviewer exactly the data it is asking about, and records who decided what and when.
human.approval
Connectors
1 of 4 run todayPost to Slack
runsPosts into a channel or thread in a connected workspace.
connector.slack.post_message
Send Telegram message
coming soonThe Telegram connection exists; the outbound send step does not yet.
connector.telegram.send_message
Send WhatsApp message
coming soonThe WhatsApp connection exists; the outbound send step does not yet.
connector.whatsapp.send_message
Post to Google Chat
coming soonThe Google Chat connection exists; the outbound post step does not yet.
connector.google_chat.post_message
Agents
1 of 2 run todayProcessor: application readiness
runsPilot: assess application readiness, propose an operations Slack follow-up for approval, and resume as the application changes. Requires workspace setup.
agent_step.core_agent
Hand off to agent
coming soonPasses a task to a deployed agent — waiting on hosted deployment itself.
agent_step.hosted_agent
Application-event triggers — created, qualified, certified — are available too. Capability nodes appear in the palette for all seven registered capabilities; five of them return a contract-locked pending envelope, which the builder tells you before you wire one in.
Where a person fits
Approval is a node, not a convention
The point of drawing the procedure is that the handoff to a human is part of it. An approval node stops the run, shows a named reviewer the data the decision turns on, and records the decision against the run.
It stops, it does not guess
Approved, rejected, and timed out are three separate branches. A flow that wants to act on a rejection has to say so on the canvas.
The reviewer sees the inputs
The node carries the data it is asking about, so approval is not a rubber stamp on a summary somebody else wrote.
Who decided is on the run
User, timestamp, choice, and comment land on the run record next to everything the flow did before and after.
What the agent knows before it acts — the knowledge layer.