Autonomous agent
Optimized for delegation, but brittle when scope, evidence, or judgment changes mid-flight.
The operator layer for agentic work
Operator-led AI workflows for work that cannot be fully automated.
operators.sh defines and runs structured human-agent workflows where a capable operator steers expert AI workers through real business work, with checkpoints, evidence, escalation, and handoff built in.
The operator is the missing layer in agentic work.
Not autonomous agents. Not manual outsourcing. Not a passive human approval gate. Operator-led work is a repeatable operating model for the messy middle where humans and expert agents succeed together.
Operator-led workflow
A trained human operator drives agent workers through defined checkpoints, audits, and handoff paths.
Human-in-the-loop
Often treats the human as an approver after the fact, not the active driver of the workflow.
OperatorSpec v0.1
A vendor-neutral standard for defining operator-led AI workflows in YAML, backed by a JSON Schema and an RFC process.
Operator Contract
What the human drives, decides, verifies, and escalates.
Agent Worker Contract
Runtimes, tools, skills, permissions, failures, and validation.
Workflow Contract
Goal, inputs, outputs, start conditions, checkpoints, and done states.
State and Memory
Durable files that preserve context outside transient chat.
Evidence and Audit
Logs, approvals, tests, source links, and residual risk records.
Handoff Contract
Managed operation, internal training, repository transfer, or archive.
Starter workflows for SMB operators.
The first public examples are practical business workflows that show the category without exposing private customer work.
01
Small business website buildout
From business intake to researched copy, implementation, QA, deployment approval, and handoff.
02
Content research production
Source capture, claim mapping, editorial drafting, review boundaries, and publishing checklist.
03
Customer support back office
Case summary, policy matching, response drafting, approval, escalation, and final disposition.
Managed operator-led workflows for businesses.
We design the workflow, source or train the operator, configure the agent workers, run the operation, and hand it over when the customer is ready.
Workflow Discovery Sprint
Research the business need, tooling, access, risks, evidence, and escalation boundaries.
Workflow Buildout
Create the OperatorSpec definition, runbooks, agent setup, templates, and validation path.
Managed Operation
Run the workflow with trained human operators and expert agent workers for as long as needed.
Operator Training
Train an internal operator to take over with the same state, evidence, and decision model.
Handoff Package
Transfer the workflow repository, docs, state, audit trail, and open risks.
A workflow is private by default, portable by design.
Customer workflows stay private. Public examples are sanitized starters that help the community learn the pattern, train operators, and adapt the standard.
specVersion: operatorspec.io/v0.1
kind: OperatorWorkflow
metadata:
name: small-business-website-buildout
operator:
role: business-operator
agentWorkers:
- name: website-builder
workflow:
checkpoints:
- Intake facts confirmed
- Mobile and desktop QA recorded
evidence:
required:
- Screenshots
- Link checks
handoff:
modes:
- managed-operation
- repository-transferLaunch playbook
The first month turns the thesis into a public standard, a useful workflow library, and a commercial offer.
Category and spec seed
Landing page, OperatorSpec repo, SPEC.md v0.1, manifesto, and RFC process.
First starter workflow
Publish the website buildout workflow with state files, checkpoints, evidence, and handoff.
Commercial offer
Services page, discovery sprint, sample SOW, intake path, and pricing hypothesis.
Community and credibility
Discussions, roadmap, submit-a-workflow template, and first good workflow issues.
Build the standard in public. Run customer work privately.
OperatorSpec should be open enough for anyone to use and strict enough to make real work auditable. operators.sh exists to operate that model for businesses that want the outcome without building the whole practice themselves.
Start with a workflow discovery sprint.
Bring a recurring business workflow that is too judgment-heavy to automate blindly and too specialized for a generalist to run alone.