Dreamer: Implementation Notes (V3)
March 22, 2026
Architecture Observations
Natural Language Compilation
Sidekick appears to use a multi-stage pipeline:
- Intent Parsing: Extract goals, constraints, tools from description
- Plan Generation: Create operational graph (nodes = actions, edges = dependencies)
- Tool Binding: Match described operations to available tool APIs
- Runtime Generation: Produce executable agent runtime
Tool Ecosystem Strategy
The $10K prize program is early-stage platform strategy. Key metrics for Dreamer:
- Tool catalog size at month 6 (target: 100+ integrations)
- Tool quality scores (user retention per tool)
- Agent complexity ceiling (how many tools per agent)
Competitive Implications
For Claude Desktop
Local-first has security advantages but limits cloud sync. If Dreamer nails cross-device state, Claude Desktop needs local sync parity or cedes mobile/second-device use cases.
For OpenAI Operator
Pre-built agents vs. user-built agents is a fundamental UX divergence. Operator optimizes for "it just works." Dreamer optimizes for "it works how I want." Both valid, different markets.
Integration Opportunities
If building tools for Dreamer:
- Target high-frequency workflows (email triage, calendar scheduling, doc drafts)
- Design for partial automation (human approval gates increase trust)
- Expose state externally (users want to see what their agent did)
- Handle ambiguity gracefully (natural language is lossy)
Questions for Builders
- What API surface does Dreamer expose for tool integration?
- How does authentication work for third-party tools?
- What's the sandbox model for agent execution?
- Can agents be exported/migrated between platforms?
Related Listening: Latent.Space full episode on Dreamer architecture and waitlist strategy.