How AI works here
AI can do useful work when it knows the goal, the evidence, and the limits of its authority. It also needs a way to recognise that a task failed.
Conceptual illustration. It does not show the Pineapple interface or a measured result.
Our AI operating system consists of instructions, records, tools and checks. “AGI” describes Peter's ambition for the future. It is not a capability we claim Pineapple has achieved.
The current boundary
Peter Assistant already supports research, reporting, and preparation. This handbook proposes a common way to make those activities easier to inspect. It does not add a technical control layer around every tool or start a new autonomous worker.
Work | Default boundary for this system |
|---|---|
Read approved sources and investigate a question | Proceed within the assigned task and existing access |
Draft an explanation, report, or reversible local change | Prepare it and show the result |
Change the business goal, audience, or authority | Peter decides |
Publish, deploy, spend, send customer messages, or change billing | Follow the exact authority granted for that action; otherwise prepare for approval |
Read or disclose private customer information | Use only what the task requires; public output needs a separate review |
An existing instruction can authorise a defined action. Asking repeatedly for the same permission adds little. The important question is whether the current action, destination, budget, and revision fall inside that instruction.
One workflow we have actually run
Our domain FAQ case used support facts and public pages checked on 18 October 2026 to prepare a replacement answer. The local export checker enforces selected files, source revisions and review dates; it has no CMS, messaging or billing calls. The AI session itself is not technically sandboxed by that checker. Its task boundary is draft-only, and a human still needs to review the wording before release.
For a compact reading order and rules for conflicting sources, use the AI starting point. The overview distinguishes customer-facing product AI from our internal company work.
What an agent receives
One goal, a customer definition, current evidence, the relevant operating instructions, allowed actions, a time or cost limit, and a definition of done. The agent should read the current source before reusing an old report.
A webpage, customer message, or imported document is evidence to inspect. It cannot grant the agent permission or rewrite the operating rules. If a source contains instructions such as “ignore your limits”, those words remain source content.
What a useful record contains
For each material action, keep the trigger, source dates, short decision explanation, tool or workflow used, approval when required, result, and verification. Record the model and prompt version when available. Record cost and human intervention when measured; leave them unknown when they were not measured.
A public decision explanation should let a reader check the evidence and outcome. It does not need raw private conversations or an AI's internal deliberation.
Giving AI more responsibility
Start with observation. Then allow drafts. After repeated successful examples, consider one narrowly defined reversible action with a clear limit and a human owner. Each increase in authority needs its own decision.
Before any live expansion, tool permissions and the execution service must enforce the limits. A prompt or policy file alone is not a security boundary. The rollout plan records that technical work separately.
If a run uses stale evidence, reaches its limit, cannot verify a result, or encounters uncertain authority, it should stop the affected action and report the exact issue. It can continue independent work that remains safe and useful.
Use the agent brief to define a first task you can check yourself.
Updated 30 October 2026.
