Skip to main content
Observerbee’s agents study your product’s state graph and goals and turn what they find into insights, design proposals, and concrete actions such as drafted tickets and pull requests. Governance controls how far each agent can go on its own; everything beyond that waits for a human.

​
Agents are organization members

Agents live in your organization alongside human members. Each agent has a persona that shapes what it looks for and how it communicates:
  • Executive: high-level trends and business impact.
  • PM: funnels, drop-offs, and prioritization.
  • Researcher: user behavior patterns and evidence gathering.
  • Engineer: technical friction and concrete fixes, up to proposed code changes.

​
Impact tiers and autonomy

Every action an agent can take is classified by impact tier, from least to most consequential: Each agent has an autonomy level on the same scale. An agent may perform actions at or below its autonomy on its own. Anything above creates an approval request instead: the action is fully prepared, then parked in the Actions queue in the dashboard until a human approves or rejects it.

​
Policies

Organizations can attach policies, expressed as JSON-Logic rules, that are evaluated against every proposed action. Policies can require approval, block, or allow based on the action’s tier, the agent, the target, and context. Policy evaluation runs regardless of autonomy level, so you can tighten the rules beyond the tier system.

​
Accountability

Every execution records an accountability chain: which agent proposed the action, under what autonomy, which policies were evaluated, who approved it, and what was executed. Nothing an agent does is untraceable.

​
Credentials never reach agents

Agents never see your OAuth tokens for GitHub, Linear, or Slack. When an approved action needs an integration, the agent receives a short-lived capability token (valid for 30 minutes) scoped to that action. A credential proxy exchanges it server-side, injecting the real token, which is stored AES-GCM encrypted and never leaves the proxy.

​
Code changes are always pull requests

When an engineer agent proposes a code change, the work runs in an isolated sandbox and the result always arrives as a pull request in your repository. Agents cannot push to your branches directly; a human reviews and merges, same as any other contributor.