Tools, actions & environment

How agents interact with the world — the core loop in detail

The environment an agent works in

Every agent operates within an environment: the set of information it can observe and the set of actions it can take. For Claude acting as a work agent, the environment is your digital workspace — email, calendar, files, web, databases, APIs.

Categories of agent actions

Read actions (observation)

  • Read an email thread
  • Search your calendar for conflicts
  • Look up a record in a database
  • Browse a webpage

Write actions (changes the world)

  • Create a draft email
  • Add a calendar event
  • Create a document in Drive
  • Insert a database record

Communication actions (outbound)

  • Send an email (not just draft — actually send)
  • Post to Slack
  • Create a task in a project management tool

Reversibility matters

Read actions: Low risk. The agent reading your email doesn’t change anything.
Write actions: Medium risk. Creating a draft can be undone.
Communication actions: High risk. Sending an email is irreversible. Design human checkpoints before these.

The action-observation cycle in practice

Example: agent handling a client inquiry email

  1. Observe: Read the email — it’s asking about project status
  2. Plan: Need to check the project file in Drive for current status
  3. Act: Read the project status document in Drive
  4. Observe: Project is 70% complete, on track for the deadline
  5. Plan: Draft a reply with the status update
  6. Act: Create a draft reply in Gmail
  7. Report: “I drafted a reply to Scott’s status email — ready for your review before sending”

Map your workflow: Think of a recurring task you handle. Map out the observe/plan/act steps. Where are the irreversible actions? Those are your human-checkpoint candidates.