01 · OBSERVE
It reads what just happened
A message, a file, an error, search results. This is the agent's only window onto the world — it knows nothing you haven't shown it.
How our agents work
Agents work in a loop: observe, think, act — then loop again.
Strip away the hype and an "agent" is a program stuck in a simple, repeating rhythm. It checks the current state of things, reasons about what to do next, takes one action, and then checks again. Below is that loop broken down, followed by three real examples doing very different jobs.
02 · THINK
It decides what to do next
It reasons about what that observation means and picks a next move: which tool to call, what to search, what to write.
03 · ACT
It does that one thing
Then it hands the result back to itself as a new observation. The loop closes and starts again, until the task is done.
Same loop, different jobs
Three agents, one rhythm
Research assistant
"What changed in the new tax bracket?"
Observereads the question, notices it needs current, sourced facts.
Thinkdecides what to search and which sources to compare.
Actwrites a short, sourced answer.
Debugging agent
"Why is checkout failing?"
Observereads the error log, reproduces the failure.
Thinkforms a hypothesis, picks the file most likely at fault.
Actedits the code, reruns the tests.
Trip planner
"Plan a weekend in Lisbon"
Observechecks the calendar and stated budget.
Thinkcompares flights and places to stay against that budget.
Actdrafts an itinerary to review.
Different tools, different tasks — every agent above is running the same loop underneath.
Ready to put an agent to work in your product?
Tell us about the workflow you'd hand off — we'll design an agent around it, not the other way around.
