Documentation

Tenant OwnerReviewerOperator

Product knowledge base, playbooks, and embedded guidance for every persona.

Agent Behaviors

CoffeeBreak agents behave predictably because they operate within a shared contract: they take structured context as input, produce a proposed result as output, and follow workspace rules around safety, review, and confidence.

This page explains what “agent behavior” means from a workspace and developer perspective.

How agents adapt over time

CoffeeBreak gets better at working with your workspace as you create tasks and review outcomes. It primarily adapts using three signals:

  1. Review decisions and feedback
    Approvals, change requests, and reviewer comments teach the system what “good” looks like for your team.

  2. Observed agent success
    When an agent consistently performs well for a type of work, CoffeeBreak will prefer it for similar tasks.

  3. Workspace configuration
    Confidence thresholds, reviewer groups, and integration rules determine what can run unattended vs. what must pause for review.

Adaptation is automatic — you don’t need to manually tune it.

What you’ll notice in practice

Over time, teams typically see:

  • Better agent selection: routing improves for your codebase and conventions.
  • Fewer avoidable review loops: repeated pitfalls show up less often.
  • Faster turnaround: more tasks land cleanly on the first pass.
  • More consistent output: results better match the patterns your team expects.

What controls agent behavior

Agent behavior is constrained by the workspace and the task. Common controls include:

  • Workspace policies (standards, conventions, allowed operations)
  • Confidence thresholds (when to require human review)
  • Reviewer requirements (who must approve)
  • Task descriptions and acceptance criteria
  • Branch protection and pull request rules

CoffeeBreak is designed to respect these constraints. If your workspace requires human approval, the system will pause rather than pushing changes through silently.

Inputs to an agent

An agent typically receives:

  • The task type and task description
  • Relevant source code and/or documents
  • Capability hints describing the skills or domain needed
  • Prior outputs and reviewer feedback (for refinement or retry scenarios)

Outputs from an agent

Agents return:

  • A proposed result (code changes, documentation, analysis, etc.)
  • A confidence score indicating how reliable the result is
  • Optional metadata (warnings, assumptions, findings, suggested next steps)

Behavior guarantees and constraints

To support retries and orchestration across many agents, CoffeeBreak expects:

  • Idempotence where possible: retries should not create compounding side effects.
  • Policy compliance: agents follow workspace rules and guardrails.
  • HITL deference: if review is required, agents do not attempt to bypass it.
  • Predictable contracts: standardized inputs/outputs so the platform can route, evaluate, and refine consistently.

Marketplace and custom agents

If you install marketplace agents or build your own:

  • CoffeeBreak will include them in agent selection.
  • Strong performers will be used more often for matching tasks.
  • Poor performers will be deprioritized over time.

This allows teams to extend CoffeeBreak without needing to hand-tune routing logic.

Privacy and safety

Adaptation is scoped to your workspace. CoffeeBreak does not share your agent preferences, review history, or feedback signals across other customers or organizations.

For developers

If you’re building a custom agent or integration, see: