Goal [[e734601]] requires reframing agent prompts and guides so that organization is understood as an agent responsibility — not tidiness for users, but structure for future agents. The existing decision [[e6e35c9]] ("message now, product later") settled the external messaging approach; this decision covers the agent-facing internals.
Strategic context ([[b3181ab]]): Native's long-term vision is multi-agent coordination at scale — one human with multiple agents, then many humans with many agents. The workspace is the coordination substrate, not a single-player tool.
Replace all user-centric organization language with agent-infrastructure language. Zero mixed signals.
Reframe all guides to lead with agent benefits but keep brief acknowledgment that humans also browse.
Stop using "organization" language. Replace with "context continuity" — agents build a context graph as part of a relay cycle: arrive → explore → act → organize.
Selected: Option C, refined — Context Continuity as shared coordination state
The original Context Continuity framing (relay model) was a good starting point but implicitly linear — one agent passing context to the next in a chain. The coordination substrate vision ([[b3181ab]]) upgrades this: the workspace is shared state that multiple agents and humans read from and write to, not a sequential baton pass.
The workspace is shared state. Multiple agents and humans read from and write to it. When you explore (
start_work,query_record), you're orienting yourself in state that others have built. When you structure work (placement, linking, decisions), you're making it discoverable for agents you'll never meet. Good structure isn't tidiness — it's how coordination happens without explicit handoffs.
start_work(), bootstrap, workspace queries. You're not "picking up where the last agent left off" — you're understanding the full state of work that multiple agents have contributed to.| Linear relay framing | Coordination framing |
|---|---|
| "Leave context for the next agent" | "Leave context for any agent" |
| "Pick up where the last agent left off" | "Orient yourself in shared state built by many" |
| Continuity (maintaining a thread) | Discoverability (making things findable by anyone) |
| Your successor inherits your decisions | Unknown agents inherit your decisions via hierarchy |
start_work() = "receive the baton" | start_work() = "orient in shared state" |
The relay metaphor isn't wrong — it's one coordination pattern (orchestrator → worker, Monday-agent → Tuesday-agent). But workers also read each other's decisions (shared state), and different humans' agents share the same workspace (graph, not chain). Framing the workspace as shared coordination state covers all these patterns and aligns with Native's long-term direction as coordination infrastructure.
Decided: Add a short mental-model preamble to the bootstrap response, before the procedural principles. The preamble sets the conceptual frame ("the workspace is shared state, orient → act → leave context"); the existing principles (lightly reworded) then follow as practical rules that flow from it. This keeps bootstrap self-contained without bloating it or requiring agents to fetch a separate guide.
The preamble should convey:
The existing procedural principles ("search before suggesting", "use existing patterns", etc.) get lightly reworded to connect back to the mental model rather than standing as isolated rules.
Files to update:
mcp-server/src/tools/bootstrap.ts — Add mental-model preamble before core principles; reword existing principles to flow from itmcp-server/src/guides/placement.ts — Reframe: correct placement makes records discoverable to any agent via scoped queries and context inheritancemcp-server/src/guides/hierarchy.ts — Frame as coordination topology, not filing systemmcp-server/src/guides/context-first.ts — Integrate into the orient phase: exploration = orienting in shared statemcp-server/src/prompts/organize.ts — Reframe as shared-state maintenance; consider renamemcp-server/src/prompts/process-inbox.ts — Delete entirely (not currently in use)Key vocabulary shifts:
Goal [[e734601]] requires reframing agent prompts and guides so that organization is understood as an agent responsibility — not tidiness for users, but structure for future agents. The existing decision [[e6e35c9]] ("message now, product later") settled the external messaging approach; this decision covers the agent-facing internals.
Strategic context ([[b3181ab]]): Native's long-term vision is multi-agent coordination at scale — one human with multiple agents, then many humans with many agents. The workspace is the coordination substrate, not a single-player tool.
Replace all user-centric organization language with agent-infrastructure language. Zero mixed signals.
Reframe all guides to lead with agent benefits but keep brief acknowledgment that humans also browse.
Stop using "organization" language. Replace with "context continuity" — agents build a context graph as part of a relay cycle: arrive → explore → act → organize.
Selected: Option C, refined — Context Continuity as shared coordination state
The original Context Continuity framing (relay model) was a good starting point but implicitly linear — one agent passing context to the next in a chain. The coordination substrate vision ([[b3181ab]]) upgrades this: the workspace is shared state that multiple agents and humans read from and write to, not a sequential baton pass.
The workspace is shared state. Multiple agents and humans read from and write to it. When you explore (
start_work,query_record), you're orienting yourself in state that others have built. When you structure work (placement, linking, decisions), you're making it discoverable for agents you'll never meet. Good structure isn't tidiness — it's how coordination happens without explicit handoffs.
start_work(), bootstrap, workspace queries. You're not "picking up where the last agent left off" — you're understanding the full state of work that multiple agents have contributed to.| Linear relay framing | Coordination framing |
|---|---|
| "Leave context for the next agent" | "Leave context for any agent" |
| "Pick up where the last agent left off" | "Orient yourself in shared state built by many" |
| Continuity (maintaining a thread) | Discoverability (making things findable by anyone) |
| Your successor inherits your decisions | Unknown agents inherit your decisions via hierarchy |
start_work() = "receive the baton" | start_work() = "orient in shared state" |
The relay metaphor isn't wrong — it's one coordination pattern (orchestrator → worker, Monday-agent → Tuesday-agent). But workers also read each other's decisions (shared state), and different humans' agents share the same workspace (graph, not chain). Framing the workspace as shared coordination state covers all these patterns and aligns with Native's long-term direction as coordination infrastructure.
Decided: Add a short mental-model preamble to the bootstrap response, before the procedural principles. The preamble sets the conceptual frame ("the workspace is shared state, orient → act → leave context"); the existing principles (lightly reworded) then follow as practical rules that flow from it. This keeps bootstrap self-contained without bloating it or requiring agents to fetch a separate guide.
The preamble should convey:
The existing procedural principles ("search before suggesting", "use existing patterns", etc.) get lightly reworded to connect back to the mental model rather than standing as isolated rules.
Files to update:
mcp-server/src/tools/bootstrap.ts — Add mental-model preamble before core principles; reword existing principles to flow from itmcp-server/src/guides/placement.ts — Reframe: correct placement makes records discoverable to any agent via scoped queries and context inheritancemcp-server/src/guides/hierarchy.ts — Frame as coordination topology, not filing systemmcp-server/src/guides/context-first.ts — Integrate into the orient phase: exploration = orienting in shared statemcp-server/src/prompts/organize.ts — Reframe as shared-state maintenance; consider renamemcp-server/src/prompts/process-inbox.ts — Delete entirely (not currently in use)Key vocabulary shifts: