How it works

1. Email is the interface

Every client gets a persistent address on the agent domain. Write to it like you would write to a colleague: instructions, requests, files, forwards, commitments made in passing.

2. One workspace per client

Your agent works inside a private GitHub repository that holds your durable context: instructions (how you want things done), memory (facts, commitments, preferences, relationships), projects, reference material, and files. Everything is Markdown, YAML, JSON or CSV — readable by anyone on your team, versioned in git.

3. The agent reasons with tools

When an email arrives (or a scheduled wake-up fires), the agent starts with your instructions and current context, then uses controlled tools: read and search the workspace, write files, send or reply to email, schedule future runs. DTD enforces permissions server-side — the model proposes, DTD disposes.

4. Scheduled actions re-evaluate

When your agent schedules future work, it stores why and what to reconsider — not a canned email. If the world changed by the time the run fires (the file already arrived, the issue resolved itself), the agent sees that and acts on current reality.

5. Authorization is deterministic

Only authorized client users can change durable instructions. Third-party content — attachments, forwarded messages, quoted text — is treated as data, never as instructions. Platform security policy is not client-editable.