“Did anyone follow it?”
The wiki says a manager approves. Nothing shows whether one did.
An AI agent SOP, or agent operating procedure, is a standard operating procedure written as workflow steps the agent executes one at a time, with stops where a person decides. ConvOps turns the SOP your team keeps in a wiki into a process the engine holds, so every run follows it and edits reach the next run.
claude · codex · cursor · any mcp client · one step at a time · a person decides at the gate
new chat
Vendor onboarding · example
The procedure lives in a wiki. People skim it, agents never see it, and the steps drift.
“Did anyone follow it?”
The wiki says a manager approves. Nothing shows whether one did.
“Which version is current?”
Three copies of the procedure, edited by three people, in three places.
“Does the agent know?”
Paste the SOP into a prompt and the model decides which parts matter today.
An example procedure, turned into a workflow from Claude Code, then run on a real case.
You hand the agent the wiki page. It proposes a workflow: one step per action, a gate where a person decides, a step that must record its reasoning.
The same procedure, now held by the engine. The agent sees one step at a time, so it cannot jump ahead to the part it likes.
The agent runs the checks the SOP names and writes the findings down. The step closes only with a reason recorded.
Checks
The run stops at the approval gate. The manager reads the findings and decides. Until approval is passed, the vendor is not created.
Approve this vendor?
Checks passed · contract over threshold · reasoning recorded
Every step lands in the history with its reason. Field changes on the task keep the agent label and the signed-in person apart.
agent label · implementer | person · verified from the credential
Each step is served when the run reaches it. The approval is where the SOP says a person decides.
Legal name, tax id, bank details and contact, as the SOP lists them.
Against the policy the SOP names. Findings recorded on the task.
The step cannot close without the reason written down.
decisionAn approval gate holds the run until a person passes it.
needs a personOnly after approval. The step history shows every stage.
The reasons behind each step, kept by the team and recalled when the next case arrives.
Why the threshold is where it is, recorded at the step that applied it.
why we do it
The check that caught a bad bank detail last month, so the next run looks for it.
what went wrong
The policy, the vendor sheet and the contract folder. Found once, kept.
where it lives
Policies attached to a workflow reach the agent with the step they apply to.
guidance, not a gate
The engine serves the current step only. Steps inside one workflow never run at the same time.
requires_approval refuses the advance until approval is passed, and the response lists what is missing.
require_context keeps a step open until the reason is recorded, so the decision travels with the work.
Step history with reasons, an activity feed, and field-level before and after on tasks and workflow instances.
Connect the AI tool your team already uses, install the workflows, and run one real piece of work through them.
Pick one SOP your team repeats weekly and paste the wiki page into your AI client
Ask the agent to draft it as a workflow: one step per action, a gate where a person decides
Run one real case through it and read the step history afterwards
Fix the step that was unclear once; every later run follows the new version
A standard operating procedure written so an AI agent can execute it: ordered steps with instructions for each, and explicit stops where a person decides. Research such as SOP-Agent (arXiv:2501.09316) shows procedure-shaped guidance helps general-purpose agents; ConvOps keeps that procedure in an engine instead of a prompt.
The model then decides which parts to follow. In a workflow, the engine serves one step at a time and checks gates before the run moves, so an approval gate pauses the run until a person approves, whatever the model judged.
Usually the agent. Paste the SOP and ask for a workflow; it creates the steps and gates over MCP, and you review them. There is no canvas editor: changes are made in conversation and stored by the engine.
Edit the step once. Runs that start afterwards follow the new version, for every person and every AI client on the workspace.
If the approval step names an on_rejected target, a rejection moves the run to that step, for example back to the checks, and the step history records it. If the step has no target, the rejection is refused and the run stays at the approval.
Not inside one workflow: steps run in order. Independent pieces of work are split into child tasks that start together, and the parent waits for all of them.