glossary · The process

Workflow policy

definition

A workflow policy is a reusable rule, written once and delivered to the AI agent inside the step it is executing, so the rule is in context at the moment it applies.

in one line: A rule delivered inside the step where it applies.

in context

Where does a workflow policy fit?

Edit once. Every run hears it. In ConvOps, the rule lives in one place, is bound where it applies, and arrives with the step it governs.

edit the ruleonce
runs in flightmany
next steprule inside
AI actswith the rule
two ways to say it

What is a workflow policy, in plain words?

Same idea at two depths: the plain version, then what the engine actually does.

Rules pasted into prompts go stale as the prompts get copied. A policy lives in one place and shows up inside the step where it applies.

A policy tells the AI. It does not stop it. Stopping is the job of a gate, a required metric or a required report.

for engineers

Reusable guidance text, bound at any of nine levels from the workspace down to one step, and delivered inside the step response.

The narrowest binding wins and an exclude removes a wider one. An edit reaches every run at its next step. Policies inform. Gates, required metrics and required reports stop a step.

the longer answer

Why does a workflow policy matter?

Policies solve the distribution problem of standards: rules pasted into system prompts drift as prompts are copied and edited, and rules in a handbook are wherever the agent is not. A policy lives in one place, is bound where it applies, and arrives inside the active step's payload.

The operational consequence is instant propagation: edit the policy once and every run in flight picks it up at its next step. No template surgery, no redeploy, no memo.

Policies inform; they do not block. What stops a step is a gate, or a number or a report the step is required to produce. A governance design that is honest about that split puts rules in policies and stop points in gates and requirements, and never claims one does the other's job.

A good policy is short, specific and scoped. "Commit messages follow the conventional format", bound to the workflows that commit, beats a page of standards attached everywhere. Scope keeps each step small, so the rule the agent needs is not buried among rules that do not apply to what it is doing. Policies also give a team one place to answer which rules the agents follow: read the policy list, not every prompt.

in convops

How does a workflow policy work in ConvOps?

In ConvOps, a policy is reusable guidance text. You write it once and bind it where it applies, at any of nine levels: the workspace, a project, a goal, a program, an initiative, a schedule, a task, a workflow or a single step. A binding can also exclude a policy at a narrower level, and when the same policy is bound more than once, the narrowest binding wins. When the agent reads a step, the policies bound along that task's line arrive inside the step's response, and an edit reaches every run at its next step.

In ConvOps, policies inform the agent. What stops a step is a gate, a required metric or a required report: the advance is refused until the approval, the number or the report is there. A preview shows which policies, metrics and reports a run will get before it starts, and each policy shows every place it is bound. Policies are written and edited from your AI client over MCP; the app and the REST API read them and bind them. Because a policy is text the agent reads, it works in every client without any client-side setup.

in depth

Where can a policy be bound in ConvOps?

A policy reaches every task below the level it is bound to. Bind a rule high when it is true everywhere, and low when it is true only for one kind of work or one step. The rows below are examples from several kinds of work.

The nine binding levels, with example policies.
LevelReachesExample policy
WorkspaceEvery task in the workspaceEnterprise: never paste customer personal data into an outside tool
ProjectThe work of one projectFinance operations: amounts are quoted in the ledger currency, with the source entry
GoalWork under one goalSales: every outreach batch states the segment it targets
ProgramThe recurring runs of one programSupport: a nightly triage run escalates, it never closes a ticket
InitiativeOne committed body of workResearch: cite the primary source for every figure
ScheduleEvery run that schedule startsOperations: the weekly access review reads, it never revokes
TaskOne piece of workContent: this launch article follows the legal team's wording list
WorkflowEvery task on that workflowContent: a claims check before anything is published
StepOne step of one workflowSoftware delivery: the commit step follows the conventional format
  1. Effect: bind adds the policy; exclude removes a binding made at the same or a wider level, for the work at and below it.
  2. Mode: enforce, observe or off. For metrics and reports, enforce is what lets the requirement stop a step.
  3. Precedence: only bindings on the task's own line count, and the narrowest one wins.
in depth

How is a policy different from a prompt rule or a gate?

All three shape what an agent does, at different strengths. A prompt rule is text someone pasted into one prompt. A policy is the same kind of text, kept in one place and delivered with the step it applies to. A gate, a required metric or a required report is a check the engine makes before the work moves on. Use the lightest one that holds, and put a check wherever a rule alone is not enough.

Three ways to shape agent work.
QuestionPrompt rulePolicyGate or requirement
Where it livesInside one prompt or fileIn one place, bound where it appliesOn the workflow step
When the agent sees itWhenever that prompt is loadedInside the step it applies toWhen it tries to advance
Changing itEdit every copyEdit once; runs get it at their next stepEdit the step
Can it stop the work?NoNoYes, until the approval, number or report is there
ExampleEnterprise: a style line in one team's promptEnterprise: one wording list bound to every content workflowEnterprise: a legal approval before publish
questions

What do people ask about a workflow policy?

What is a workflow policy in ConvOps?

A workflow policy in ConvOps is a reusable rule, written once and delivered to the AI agent inside the step it is executing. It is bound where it applies, at any of nine levels from the whole workspace down to one step, and an edit reaches every run at its next step. A policy informs the agent; it does not stop a step.

How do I check what a policy will deliver to a step?

Run a guidance preview before any run reaches the step. In ConvOps the Simulator, reached from the policy pages, and POST /guidance/preview show which policies, metrics and reports a given step would receive under the current bindings, and they warn about required items that can never apply. Fix the binding scope there, rather than learning from a run that the rule never arrived.

Where can a policy be bound?

In ConvOps, a policy can be bound to the workspace, a project, a goal, a program, an initiative, a schedule, a task, a workflow or a single step. Only bindings on the task's own line count, the narrowest one wins, and an exclude removes a binding made at the same or a wider level for the work below it.

How is a policy different from a rule in CLAUDE.md?

A CLAUDE.md rule is loaded into every session whether it applies or not, and each repository keeps its own copy. A ConvOps policy lives once in the workspace, is bound only where it applies, and arrives inside the step it governs, so the agent sees it when it matters. Editing it once updates every run at its next step.

Do workflow policies work in every AI client?

Yes. In ConvOps a policy is delivered as part of the step over MCP, so every connected client receives it without any client-side setup, whether that is Claude Code, Codex, Cursor or an unattended run. Policies are written and edited from the AI client over MCP, and the web app and REST API can read them and bind them.

Who writes workflow policies?

The people who own a rule: a security lead, a brand editor, a compliance owner or a tech lead. In ConvOps they write a policy once and bind it where it applies, and runs receive it inside the step it governs, so the rule reaches each AI client without anyone editing instruction files in each repository.