State model
Define valid states, transition conditions, timeouts, and terminal outcomes.
State diagram and transition contractEnterprise AI · Agentic workflows
Give automated work a bounded state machine, explicit permissions, and accountable review.
Agentic workflows can coordinate evidence gathering and tool use, but production design must make state, authority, retries, approvals, and compensation visible. We engineer the workflow as a service rather than granting an open-ended objective.
- Boundary
- Declared tools, states, budgets, and stop conditions
- Authority
- Least privilege with approval for consequential actions
- Evidence
- Replayable decisions, tool results, and state transitions
02
01 · Workflow model
A reliable agent starts as a visible workflow with named states and owners.
We map triggers, evidence inputs, deterministic services, judgment points, side effects, failure routes, and handoffs. Open-ended goals are converted into bounded transitions that can be observed and tested.
Authority map
Tie each read, proposal, and write to an identity, permission, and approval rule.
Tool and permission matrixRecovery model
Specify retry, idempotency, compensation, pause, and human takeover behavior.
Failure-mode and recovery table03
02 · Execution
Tools need narrow contracts and policy enforcement outside generated text.
Each integration receives validated input, scoped credentials, deterministic policy checks, time and resource limits, and structured results. Consequential actions remain separated from planning.
Propose
Create a structured action proposal with evidence, expected effect, and required authority.
Schema validation and proposal traceAuthorise
Evaluate policy and obtain accountable approval before protected execution.
Policy result and approval recordExecute
Use idempotent interfaces and capture the result needed to reconcile workflow state.
Tool receipt and state transition04
03 · Evaluation
Success cases are the smallest part of agent readiness.
Scenario tests exercise malformed tool output, unavailable dependencies, duplicate requests, expired approvals, denied permissions, contradictory evidence, and interrupted execution.
Control cases
Verify that policy and approval cannot be bypassed by workflow state or generated output.
Denied-action and expired-authority resultsRecovery cases
Verify retry, compensation, and reconciliation under partial side effects.
Fault-injection and recovery recordHuman cases
Verify that reviewers receive enough context to approve, reject, amend, or take over.
Reviewer task-completion findings05
04 · Operations experience
Operators should be able to answer what happened, why, and who can act next.
The operating surface presents state, elapsed time, source evidence, tool receipts, approvals, errors, and safe intervention commands without requiring access to raw internal logs.
Illustrative operations surface: state, intervention, and ownership remain explicit.
State timeline
Transitions and waits are visible with timestamps and accountable actors.
Intervention controls
Pause, cancel, retry, reassign, and reconcile actions follow explicit permissions.
Case bundle
A trace identifier links the user-visible workflow to retained technical evidence.
06
05 · Handover
An agentic workflow needs an owner for every state, including stuck and unsafe.
The production pack includes state and tool contracts, permission reviews, evaluation scenarios, support thresholds, recovery playbooks, and a controlled process for adding tools or authority.
Service duty
Named teams own workflow availability, business exceptions, and integration incidents.
Authority review
Credentials, permissions, and approval rules are revalidated on a defined cadence.
Change discipline
New tools and transitions require scenario evidence before release.
The next step is a bounded working session: one workflow, its evidence sources, the decision owner, and the conditions under which the system must stop or hand over.
Map an agentic workflow