Continuous reflection: the missing loop in agentic work

At a recent internal Show and Tell, Xiaozhen Zhu and I compared notes on how a year of building with AI has changed the way we work.

The biggest shift is not faster typing. We now spend less of our time typing code. More time goes into setting up the environment, explaining the goal, and deciding whether the result is good.

As the tools become faster, direction and constraints become more important.

One loop I took away from the talk is simple:

1. Design the change before asking for code.
2. Set rules and boundaries with types, tests, permissions, and other constraints.
3. Let the agent run checks and receive errors while the context is still fresh.
4. Review the result, then review the workflow itself.

That last step is what we called continuous reflection.

It is easy to settle into the comfort of a workflow that already works. As long as it gets the job done, there is little pressure to question it. But our tools are changing quickly. Familiar habits can stop us from seeing better ways to work. We risk making new agentic capabilities fit our old processes instead of rethinking the processes themselves.

Continuous reflection means stepping back regularly. Look for repetition, unnecessary handoffs, and missing feedback loops. Ask whether agents could compare options, challenge one another’s assumptions, run checks themselves, or work with better tools and context.

After an agentic session, ask:

• What did I repeat?
• What should have been automatic?
• What context, rule, or tool would improve the next session?
• What new capability am I overlooking because the current workflow feels good enough?

Improving one result helps once. Improving the workflow carries into every session after it.