Agent work continues.
Your controls stay with it.
An agent harness is the execution layer around a model: it manages task state, tool access, approvals, and limits. CapitalSource brings that layer to commercial finance, starting with the Processor application-readiness pilot.
Processor pilot · Workspace setup required · Broader fleet support planned
Task lifecycle
- 1
Start
A published flow starts a Processor task for a selected application.
- 2
Assess
Check canonical requirements and document review evidence.
- 3
Wait for approval
Persist the proposed operations follow-up for a designated reviewer.
- 4
Resume
Reassess as the application changes, preserving task history and limits.
Build on a task that keeps its context.
Durable task state
Flow versions, agent turns, proposals, and receipts are persisted. Application updates can resume the same task with its execution history intact.
Scoped execution
Each task carries the initiating user’s authority and explicit resource scope. Current membership, credentials, and connection status are rechecked during execution.
Approval bound to evidence
A reviewer approves the exact Slack text and destination. Changes to the application or selected published knowledge invalidate the previous approval.
Bounded model work
Model calls reserve budget before execution. Task spending, turn counts, and time limits bound the work; actual and reserved spending remain visible.
Uncertain deliveries stop for investigation. An owner or admin records a verified outcome before reassessment; the harness does not silently retry an ambiguous Slack send.
Put task visibility in your own product.
Read the run. Surface the next action.
The portal uses the same Core endpoints exposed by the TypeScript SDK. Read a flow run to display agent status, waiting reasons, action receipts, and model spending in your own operations interface.
This read-only example assumes an authenticated CapitalSourceClient named client and a flow run ID from your organization. Setup and run creation happen in Flows or through its API. Approval and execution controls enforce the caller’s role and task scope.
TypeScript · Inspect a flow run
const run = await client.flows.getRun(flowRunId);
for (const task of run.agent_runs ?? []) {
console.log({
status: task.status,
waitingFor: task.waiting_reason,
spentCents: task.spent_cents,
reservedCents: task.reserved_cents,
budgetCents: task.budget_cents,
});
}GET /v1/flows/runs/{runId}
Inspect agent tasks, actions, turns, and model spending.
PUT /v1/agent-harness/settings
Owner/admin control of disabled, shadow, or approved mode.
POST /v1/agent-actions/{id}/decision
Record an authorized reviewer’s decision with the exact proposal hash.
POST /v1/agent-runs/{id}/resume
Request reassessment within the original grant, budget, and lifetime.
POST /v1/agent-runs/{id}/cancel
Cancel the task and revoke further execution authority.
POST /v1/agent-actions/{id}/reconcile
Record an operator-attested resolution for uncertain Slack delivery.
MCP exposes application and capability tools; Recurser provides the existing CLI and MCP setup commands. Harness task controls currently use REST, the SDK, and the portal. Dedicated MCP tools and CLI commands for these controls are not available yet.
Prove the workflow in shadow mode.
Configure model routing and pricing, connect Slack, and select the operations channel and approvers. An owner or admin enables shadow mode; your team starts the Application readiness flow and reviews its evidence and proposed follow-ups before enabling approved delivery.
This pilot checks application readiness and prepares operations messages. Hosted-agent handoffs and autonomous delegation across the fleet remain roadmap work. Financing decisions and fund movements are outside the pilot.