Workflow App Sprint Category
AI Agent Action Gate
For AI product, automation, and operations teams that need a governed action boundary before software changes customer state.
The workflow
Where this breaks today.
Current failure mode
Agents can book, refund, cancel, qualify, message, escalate, or use tools before confidence, consent, evidence, and reversibility are clear.
Focused app surface
An action intake, guardrail rulebook, Decision Record, and execution handoff that returns proceed, block, or human review.
Runtime or record layer
Binding mode: trusted_adapter_facts_then_declarative_rulebook. Customer executable rulebooks are outside the current production contract.
Example boundary
What the app decides or hands off.
Agent wants to act with customer context, requested tool, confidence, policy, missing evidence, reversibility, consent, and escalation owner attached.
Proceed, block, or human review required before the agent changes customer state or uses the tool.
One uncontrolled agent action can create support cost, revenue leakage, or customer trust damage.
What ships
One useful workflow surface, not a broad project.
Trigger, current path, owner, failure modes, and the boundary where work should proceed, block, review, or route.
Required fields, evidence standard, missing-input states, and examples of acceptable requests.
Direct declarative Rulebook v1, trusted-adapter fact contract, or Decision Record-only path depending on the workflow.
A focused Krafthaus page, notary, packet builder, score, dashboard row, or handoff artifact the team can review.
What gets stored, where the record travels, and who or what receives the next action.
What to automate, measure, or turn into repeat runtime usage after the first workflow proves useful.
Trust boundary
Start with metadata and handoff shape.
Runs on action metadata and trusted adapter facts first. Full transcripts, private customer records, and full tool logs can stay outside the sprint.
One workflow
The sprint deliberately avoids broad AI transformation, platform rebuilds, or customer executable rulebooks.
Metadata first
The first sprint can use request metadata, policy fields, evidence status, owner, risk, and desired handoff before private-system integration.
15-minute review
The low-friction next step is to confirm the workflow owner, required inputs, Decide binding, and whether a Krafthaus surface is worth scoping.
Workflow App Sprint
Map your version of AI Agent Action Gate.
Use this category as the public pattern. The company-specific artifact comes after the workflow is real enough to review.