We may be trying too hard to tell AI agents exactly how to work. A better design question might be: what happens when a capable system gets a clear objective, useful resources, and firm boundaries?
Choose a real business objective. Add or remove what the agent can see and do. The result is a design hypothesis for you to test, not a prediction of agent performance.
With CRM history, search, and drafting available, the agent has room to gather context and prepare a next step.
Human gate: Nothing is sent. A person reviews the draft and the action log first.
The instinct makes sense. Define the process. Break it into steps. Assign the tools. Add the rules. That is how traditional software gets built. I am increasingly unconvinced it is the right model for agents.
Large models regularly demonstrate behaviors nobody finely scripted. When we connect those models to documents, databases, other agents, and business tools, we are changing the environment around capabilities that already exist.
A tool is more than another feature on a list. It can create a new path for a model to reuse what it already knows how to do. That path still needs observation, testing, and a clear stop condition.
Foundation model providers have already spent billions training and steering these systems. Coding agents add another layer. Keep piling on rigid instructions and we may squeeze out the flexibility we were hoping to use.
Decide what data exists for the agent. Choose the tools, permissions, approval thresholds, and logging. Then watch what happens, keep the safe surprises, and revise the environment. That is still engineering. It feels a lot more like managing an ecology than drawing an assembly line.
The next capability jump may come from new opportunity spaces we create around what is already here.
The Stockholm Paradigm grew from work on ecology, evolution, and emerging disease. One of its core ideas, ecological fitting, describes organisms forming new associations by using capacities shaped in earlier conditions when a changed environment presents a new opportunity.
I think that pattern gives us better language for agent design. A model connected to new data and tools may find uses nobody specified step by step. Still, this is an analogy. Biology does not prove how an AI system will act, and a simulated page like this one does not show real agent performance.
Pick a recurring job where a wrong move is recoverable. Give the agent enough room to be useful and enough structure to make the result inspectable.
Write the business result in one sentence. Avoid prescribing every click.
List the approved data, tools, and other agents that belong in this test.
Require human approval before spending, sending, publishing, deleting, or changing access.
Log the path, inspect the evidence, and keep what worked. The useful discovery becomes part of the next environment.
New working pages and practical AI experiments, sent to your inbox. The archive stays here whenever you need it.