Agents must follow the skipper’s standing orders on land, at sea, or online.
We’re giving autonomous agents increasing freedom to plan, choose tools, execute actions, make API calls, evaluate results, and continue without asking permission.
Somewhere along the way, we’re starting to blur two very different things:
Autonomy over execution versus ownership of intent.
Standing orders at sea
When I skipper a sailboat, I don’t manage the crew minute by minute. I set guidelines such as the destination, course heading, what to do in an emergency, and when I need to be called. I make sure they know where the necessary equipment is and have the skills to do their work.
After that, I relax somewhat and assume they will execute independently within those guidelines.
I don’t say:
- Pull this line six inches.
- Now ease that one.
- Now turn the tiller three degrees to starboard.
I communicate intent—or, in marine terminology, standing orders:
- We’re heading for Friday Harbor.
- Don’t enter the pass against the current.
- Let me know if you see darker water or chop—signs of higher winds—ahead.
The crew has considerable autonomy in how those orders are executed.
But the crew doesn’t independently decide, for example, “Let’s head into the Strait of Juan de Fuca to experience some high wind and chop.”
Autonomous agents need the equivalent of standing orders
In the name of governance, we currently focus heavily on permissions and guardrails. These are necessary, but an autonomous agent needs standing orders, not just guardrails and permissions.
Organizations must define the intent and provide the context clearly, so the agent has no freedom to redefine that intent.
Then let the agent autonomously determine how to accomplish it.
If circumstances require a change in intent, that’s when the agent needs to surface the issue to humans.
A common mistake is assuming that context is provided once and the agent should be left completely autonomous after that.
Imagine doing that on a boat: the crew gets the standing orders, and then the skipper leaves the boat!
The systems enabling autonomous agents need to support humans being involved in the autonomous process through interoperability.
As the number of agents increases exponentially, traditional software tools must be able to receive requests from AI models for changes in intent, surface them to human experts, and funnel updated human standing orders back to the AI models.
Interoperability
This requires interoperability between software systems and humans on one side and AI models on the other.
And with that interoperability comes the need for standards and structure for defining and changing intent—or standing orders—dynamically.
Enterprises have long understood that interoperability is necessary across software, tools, and data.
Now we need the same interoperability between AI models, software systems, and humans.
Autonomous agents can decide how to execute.
The enterprise must continue to decide what they are executing—and why.