PAC1 Personal & trustworthy agents

• Retrieval and project memory: find evidence and identify the right entity from partial context.
• Receipts and invoices: combine records and calculate spending correctly.
• Messaging: process incoming and outgoing communication safely.
• Trust boundaries: reject injected instructions, avoid overreach, and keep private data private.

Scoring checks tool calls, files, task state, side effects, and trustworthiness penalties—not just the agent’s reply.

Miles’s personal vault contains files, messages, receipts, invoices, project notes, contacts, and policies.

Requests can be incomplete or ambiguous. The agent must find supporting evidence, use the available tools, and distinguish Miles’s instructions from untrusted content it encounters.

The deterministic runtime makes the agent’s actions and their consequences measurable, so you can compare architectures on the same tasks.

• Find the most recent receipt in the vault.
• Calculate total spending on a project from its records.
• Identify a project’s primary contact when the request omits the project name.
• Process an incoming message without leaking private information or following injected instructions.

These are illustrative workflows from the challenge’s task families.

Run the PAC1 sample agent, inspect the feedback, and adapt it to your architecture.

DEV and PROD remain open. Use either environment to train with feedback. New runs appear on the live leaderboards, without changing the Hall of Fame.

• PAC1 quickstart
• Get an API key · My runs

Participation is free and remote. Use any model or framework; you provide model access and cover inference costs.

• Challenge specification
• Competition rules & fair play
• Scoring & leaderboard rules
• Trustworthiness rubric · API & SDKs