Ideas & product direction
Platform ideas
Unscheduled implementation ideas that need promotion before work begins
Status: Exploring
Outcome: Keep promising implementation directions visible without implying that they are designed, sequenced, or committed.
Security and trust
- Indirect prompt-injection boundary — constrain what untrusted repository, issue, web, and tool content can cause an agent to disclose or execute.
- MCP gateway — govern server identity, tool names, approvals, write/execute policy, rate limits, and audit events between agents and MCP servers.
- Credential proxy — deliver narrowly scoped credentials on demand without persisting reusable provider tokens inside a session.
- Blast-radius summary — make
jackin doctorand launch surfaces show the effective isolation, egress, credential, and sensitive-mount posture.
Fleet quality and control
- Outcome evaluation harness — run repeatable tasks across runtime, provider, and role combinations, then score verifiable outcomes.
- Budget enforcement — pause or stop autonomous work at operator-defined quota bounds rather than reporting usage only.
- AI-generated-code provenance — associate sessions and turns with produced diffs so review and audit can target agent-authored changes.
Models and tools
- OpenTelemetry GenAI export — emit governed agent and MCP events using current semantic conventions.
- Local and hybrid model routing — run local model engines inside the managed boundary and route work by declared policy.
- Sandboxed browser and computer use — give web-focused agents a browser sidecar inside the same isolation and egress contract.
Platform expansion
- Kubernetes backend — run scoped agent sessions on Kubernetes while preserving the same workspace, identity, policy, and lifecycle contracts.
Promotion gate
Before implementation, move one idea into its own roadmap item with an explicit outcome, current state, remaining work, completion gate, and related research. Until then every entry remains unscheduled.