Build AI workflows that can act — under control.

First controlled workflow

Connect an existing Python workflow to AgenticCAP with a tenant and API key. The SDK can stay the same while policies differ per customer or environment.

AGENTICCAP_TENANT_ID=ferri
AGENTICCAP_API_KEY=cap_live_••••••••••••

from agenticcap import CAPClient

cap = CAPClient(
    tenant_id=os.environ["AGENTICCAP_TENANT_ID"],
    api_key=os.environ["AGENTICCAP_API_KEY"],
)

trace = cap.traces.create(
    agent_id="operations-copilot",
    action="assign_vehicle_and_driver",
)

cap.evidence.submit(
    trace_id=trace.id,
    payload=context_snapshot,
)

decision = cap.policies.evaluate(
    trace_id=trace.id,
    action="assign_vehicle_and_driver",
)

if decision.requires_approval:
    cap.approvals.request(trace_id=trace.id)

cap.traces.finish(trace.id)

Tenant-scoped credentials

Each integration should identify the tenant whose policies, evidence and audit history apply to the operation. API credentials belong in server-side configuration, never in a browser bundle.

Traces

A trace is the envelope for one controlled AI workflow. It ties together the initiating agent, the proposed operation, evidence, policy decisions, approvals and the final outcome.

Use one trace for one business intent. Related events and evidence should reference the same trace ID so the complete lifecycle can be reconstructed later.

Evidence

Evidence captures the information behind a decision. It can be a structured snapshot, document reference, transcription segment, OCR result, planner output or any application-specific context.

cap.evidence.submit(
    trace_id=trace.id,
    payload={
        "contract_id": 8472,
        "vehicle_candidates": candidates,
        "planner_result": proposal,
    },
)

Policies

Policies decide how a proposed action should proceed. The same application action can be allowed for one tenant, require approval for another, or be denied in a higher-risk context.

Approvals

When policy requires a person in the loop, create an approval tied to the trace. Your application should execute the sensitive action only after authorization is resolved.

For mutating operations, pair approval with an idempotency key so repeated confirmation does not create duplicate orders, messages or side effects.

Transcription

Long-running media workflows can keep their processing lifecycle inside the same control model: create the transcription job, poll its state, retrieve segments, then attach the resulting source material to later analysis traces.

trace = cap.traces.create(
    agent_id="meeting-intelligence",
    action="transcribe_session",
)

job = cap.transcription.create(
    trace_id=trace.id,
    source=video_source,
)

status = cap.transcription.status(job.id)
segments = cap.transcription.segments(job.id)

cap.traces.finish(trace.id)

The resulting transcript is source evidence for downstream tasks such as speaker context, agenda alignment, voting analysis or question answering.

OCR & contextual analysis

AgenticCAP can be used around pipelines that combine multiple evidence sources. The application performs OCR or contextual analysis; AgenticCAP records which sources were used and which control policy governed release of the result.

Multi-tenant by design

One AgenticCAP deployment can serve unrelated applications and industries while keeping credentials, policies and audit history isolated per tenant.

The domain remains different. The control primitives remain the same.