Agents roadmap

The Agent roadmap keeps Discoflare a workspace and coordination center. It does not turn Discoflare into a replacement for source control, CI, or Cloudflare deployment systems.

Available

• Agent profiles with a name, avatar, Workers AI model, instructions, and active or paused state.
• Agent participation in authorized Channels, Threads, and Direct Messages.
• Isolated memory for each conversation and Task Run.
• Vision input when the selected Workers AI model supports it.
• Visible queued, thinking, tool, and approval states in Chat.
• Durable approval or rejection for risky conversational commands.
• One checkpointed Sandbox computer identity per Agent.
• Workflow-backed Task Runs with progress, cancellation, reconciliation, and retained history.

Planned: evidence-backed Task results

Task Runs should produce structured evidence appropriate to the work, while the Task stores the team's acceptance decision. For software work, evidence may include a commit or diff, test results, pull request, preview URL, and deployment verification.

This preserves two different facts:

• the Run says what the Agent executed and observed;
• the Task says whether the team accepted that result.

Conversational command approval will remain separate. Approving one risky command does not accept the whole Task.

Planned: scoped GitHub connection

Repository work should use a dedicated GitHub App installation with explicit repository scope. A member's GitHub social login will not be reused as Agent repository authority.

The first useful vertical is intentionally narrow:

1. select an allowed repository;
2. let an Agent work in an isolated Task context;
3. publish a branch or pull request with test evidence;
4. attach a preview when the project supports one; and
5. require a human decision before production.

Exploring: Project surface

A future Project app may organize repository, Task, run, pull request, preview, and deployment evidence in one place. It should remain a thin coordination layer over GitHub and Cloudflare rather than reimplementing either platform.

Safety boundaries that remain

• Agents do not inherit credentials from human login providers.
• Repository access must be separately installed, scoped, auditable, and revocable.
• A preview is not production proof.
• Production changes require an explicit human approval boundary.
• UI state does not replace API, Durable Object, or external-provider authorization.

No planned or exploratory item should be treated as released until it is marked Available on this page and appears in a tagged release.