Describe intent
The user explains in natural language which existing configuration to change and the desired outcome.
PRODUCT EXPERIENCEProduct experience is coming soon.
The public Lattice.Hub Console experience environment is being prepared.
POLE AGENT / HUMAN-GATED CONTROL PLANE
Help Agents understand change—without releasing for you.
Within signed-in user permissions, Pole Agent reads Namespaces, the MCP Registry, and existing configuration through Pole MCP. The LLM uses that context; internal proposal tooling and a confirmation kernel form update proposals. After human review and confirmation, it saves edit-state drafts only—release stays human-gated.
01 / REAL WORKBENCH
Safety boundaries should be visible in the workbench.
Pole Agent shows current resource scope, connection state, memory window, and operation permissions. When runtime configuration is incomplete, the workbench stays non-executable instead of letting the model proceed with unknown context.
Console · Pole Agent ReadinessReal readiness checks: fail closed when configuration is incomplete; resource queries allowed, change generation requires confirmation, direct release forbidden.
02 / CONTROLLED CHANGE
One reviewable change—not one untraceable command.
Pole Agent’s value is not bypassing the control plane—it translates natural-language intent into inspectable diffs and returns explicit human confirmation to existing product flows.
Read context
Read Namespace, MCP Registry, and configuration files within signed-in user permissions.
Form proposal
The LLM uses Pole MCP context; internal proposal tooling and the confirmation kernel produce an immutable proposal.
Inspect diff
Original and proposed content appear side by side for human judgment.
Confirm draft
After human confirmation, Agent writes content to edit-state draft only.
Enter release
Subsequent release, rollback, and delete remain in deterministic product flows.
03 / SAFETY LAYERS
Prompts express intent; boundaries must be enforced by the system.
From tool allowlists to the confirmation kernel to an independent release chain, each layer narrows what Agent may do and preserves auditable product semantics.
Pole Agent
Organizes conversation, context, and proposal preview so intent and diffs stay understandable.
Pole MCP
Exposes allowlisted tools with explicit parameters—does not hand arbitrary internal capability to the model.
Confirmation kernel
Validates signed-in identity, resource permissions, proposal content, and confirmation action—prompts are not the security boundary.
Product release chain
Activation, canary, rollback, and delete beyond drafts still enter deterministic human flows.
04 / THREE ROLES
Assistant, tool catalog, and Agent catalog—each owns one job.
Three concepts serve an AI-native control plane but do not replace one another. Separating roles prevents mistaking “discoverable” for “automatically executable.”
Pole Agent
Understands control-plane context, prepares change, explains diffs—does not perform final release for the operator.
MCP Registry
Registers discoverable MCP Servers, tools, and capability metadata—does not execute arbitrary tools.
A2A Registry
Registers Agent Cards, skills, and capability metadata—does not run Agent tasks or host runtimes.
05 / CURRENT BOUNDARY
State clearly what works today—and what does not yet.
The site describes only the minimal closed loop that exists now. Boundaries will evolve with the product—we will not substitute roadmap for current fact.
Closed loop in place
- Read Namespace, MCP Registry, and existing configuration files
- LLM with Pole MCP read-only context; internal proposal tooling forms update proposals for existing configuration
- Show immutable proposals and diff previews
- Save edit-state drafts after human confirmation
Out of scope today
- Automatic release, rollback, or delete
- Governance rule writes and new configuration files
- Generic writes to arbitrary control-plane resources
- Streaming output and server-side session persistence
06 / HUMAN RELEASE CHAIN
Agent prepares change; humans own release decisions.
This is not weakening automation—it keeps AI in production control planes subject to permissions, versions, release, and rollback semantics.
01Natural-language intent
02Immutable proposal
03Diff preview
04Human confirmation
05Edit-state draft
06Deterministic release
HUMAN-GATED BY DESIGN