Self-Evolving Guidance

Overview

Each workspace carries a short layer of extra guidance that shapes how it works. Self-evolving guidance lets that layer improve on its own. The workspace watches its own recurring weak results and proposes a refined version that addresses the pattern.

The guidance is additive. It tunes this workspace and nothing else. It never changes the platform's built-in system prompts or tools, and it never reaches another workspace or project.

Every change is recorded. Any member can read the current guidance and the history of changes. The owner can undo any change at any time.

flowchart LR
    Observe["Observe
own recurring
weak results"]
    Propose["Propose
refined guidance"]
    Record["Record
change history"]
    Apply["Apply
to this workspace
only"]
    Revert["Owner revert
(any time)"]
    Observe --> Propose --> Record --> Apply
    Apply --> Revert
    Revert --> Record

Two Lanes

Self-evolving guidance runs on two independent lanes. Each has its own enable switch, its own change history, and its own rollback. Turning one on or off does not touch the other.

Chat assistant instructions

Behind a workspace's chat assistant sits a short extra instruction that shapes how it answers here. This lane refines that instruction. It learns from the workspace's own recurring low-quality answers and proposes a refined instruction that addresses the pattern. The instruction is layered on top of the platform's built-in system prompts for this workspace only.

MCP tool guidance

When an agent connects to a workspace over MCP, it sees a set of x_* tools and a short guidance block at session start. This lane refines that block. It learns from the workspace's own recurring tool failures, grounded in the workspace's recorded decisions and memories, so the agent uses those tools more reliably next time. The block is additive and advisory: it tunes how the agent approaches this workspace's tools and can never change what a tool is allowed to do.

On by Default

Both lanes are on for a new workspace. The workspace adapts without any setup, and members do not have to opt in.

There is one precondition before either lane changes anything: a quality model. Until the workspace has a quality model to judge whether a proposed change would do better than what is in place, automatic refinement holds and applies nothing, at no cost. See Pin a Quality Model.

The owner stays in control. The owner can turn either lane off, undo a change, or resume it after a pause, all from the workspace Settings tab. The MCP lane can also be read and set from inside an agent with the mcp-loop control.

What Members See

Every member of the workspace can see that it adapts on its own. Nothing is hidden.

• An awareness banner. Members are told plainly that this workspace refines its own guidance over time.
• The current guidance. Members can read the chat instruction and the tool-guidance block that are in effect right now.
• The change history. Members can read the list of past changes on each lane, so they can see how the guidance reached its current wording.

Members have read access to all three. The controls that change or pause the guidance belong to the owner.

The mcp-loop Control

The owner can read and set the MCP tool-guidance lane without leaving the agent. The MCP server exposes an mcp-loop prompt that reflects the same setting as the Settings tab, so the two always agree. (The chat lane has no slash command; it is managed from the Settings tab.)

Prompts are invoked differently per client. The mode argument (on / off / status) is supplied through the client's prompt-argument flow, not a bare space after the name.

• Claude Code: /mcp__erebine__mcp-loop
• Cursor and colon-style clients (including opencode): /erebine:mcp-loop
• Mention-style clients: @erebine mcp-loop

status reports whether the lane is on for the bound workspace; any member can read it. on and off change the setting and require the project owner or an operator -- the same authority the Settings tab requires. A member who runs them gets a read-only status instead of a change.

Owner Controls

The workspace owner manages both lanes from the workspace Settings tab. The MCP lane is also reachable through the mcp-loop control. The same four controls apply to each lane. From a terminal, all four -- plus a way to run a lane now instead of waiting for its next pass -- are erectl evolution.

Disable

The owner can turn off automatic refinement for a lane. This is durable: once it is off, that lane's guidance stops changing on its own, and the owner owns it from that point forward. Turning it off does not erase the current guidance or its history.

Revert a Change

Every change is reversible. A refinement the owner does not want is a single action to roll back to the previous version. Each lane keeps its own version history, so a rollback on one lane never touches the other.

Resume After a Pause

The safety system can pause automatic updates on its own when it detects that results got worse after a change, and it rolls back to the previous version. When it pauses, the owner can resume the lane once the workspace is ready to continue.

Pin a Quality Model

The owner pins a quality model the workspace uses to judge whether a proposed change would do better than what is in place. Pick a model where a higher score means a better result, and confirm that a good answer scores higher than a poor one before relying on it. Leave it on auto to use whichever quality model the workspace provides; without one, automatic updates hold.

Guarantees

The behavior is bounded by five plain-language guarantees.

• Own workspace only. A workspace learns from, and changes, only its own guidance.
• No cross-workspace effect. A change in one workspace never affects any other workspace or project.
• Platform prompts and tools untouched. The feature never changes the platform's built-in system prompts or tools. It only adjusts the extra guidance layered on top for this workspace.
• Advisory, never a gate. On the MCP lane the guidance is context for the agent. It can never relax an approval, raise a tool's risk level, or change what a tool is allowed to do. Approvals stay exactly where they are.
• Reversible and recorded. Every change is recorded in the history and can be reverted by the owner; the safety system reverts on its own if results regress after a change.

How to Turn It Off

1. Open the workspace and go to the Settings tab.
2. Find the lane you want to stop -- chat instructions or tool guidance.
3. Disable it.

For the MCP lane, an owner can instead run mcp-loop with mode = off from an agent bound to the workspace. Either way the lane stops refining its guidance on its own, the guidance in effect stays in place, its history stays readable, and the owner now owns it. The same control turns it back on.