Skip to content

SOP Optimizer Biz Org+ ​

The SOP Optimizer is a self-optimizing prompt compiler that analyzes historical policy violations and trace logs to automatically refine and optimize SOP instructions.

What you'll learn ​

  • How the optimizer analyzes underperforming SOP clauses
  • How to trigger manual optimization cycles
  • How to review and approve optimization proposals
  • Feature gating and role requirements

INFO

The SOP Optimizer requires the feature.metaclaw flag to be enabled for your workspace. This feature is available on Biz Org, Enterprise and Self-host, and during a trial.


How the SOP Optimizer Works ​

The optimizer continuously monitors your SOP enforcement data and identifies rules that aren't performing as intended. It analyzes three key signals:

  • Enforcement patterns — How often a rule triggers (overly active vs. stale).
  • Compliance scores — The rate at which developer agent sessions violate or bypass specific rules.
  • Trace logs — The underlying token prompts and tool calls.

An hourly background sweep runs an evaluation cycle for every eligible workspace automatically. This scheduled path is proposal-only — it always generates candidate proposals for human review, regardless of your workspace's auto-apply setting. Only a manually triggered cycle (below) can auto-apply a proposal, and only when your workspace has auto-apply enabled.


Accessing the Optimizer ​

Unlike general SOP listings, the optimizer has a dedicated home: Labs › SOP Optimizer (/labs/sop-optimizer; Owners, Admins and Engineering Managers, on a plan that includes it).

1. Manual Trigger ​

While the optimizer runs hourly compiles in the background, you can trigger an immediate evaluation cycle by clicking Run optimizer. Owners and Admins also have Mine incidents, which drafts amendments from recent incidents.

  • Rate limits: Trigger is restricted to once per hour per workspace, independently of the background sweep's own hourly cadence.
  • Report: On complete, displays candidates evaluated, proposals created, and runs skipped.

Dashboards ​

The SOP Optimizer page has two tabs. Pending Proposals (with the count of proposals waiting) opens first; Run History is the second.

Run History ​

Lists all completed optimization runs:

  • Guideline — The targeted SOP and the clause index of the markdown instruction section under audit.
  • Simulated Compliance — Before/after comparison showing the change in compliance score if the proposal is adopted (e.g. 70% → 95%).
  • Delta — The net compliance improvement (runs with a delta < 5% are skipped).
  • Sample Size — Number of traces evaluated in the simulation.
  • Score — Semantic alignment rating of the mutated rule.
  • Status — Displays whether the outcome was Accepted or Rejected.

Pending Proposals ​

Lists all mutations awaiting review. Proposal source filters them: All, LLM-evolved, Incident-mined or Health scan.

  • Proposed title and score — With a badge for its source.
  • Why, and how it scored — Plain-text explanation of why the rule mutation was recommended. Incident-mined and health-scan proposals also show their incident count and threat score.
  • Clause Diff — A side-by-side comparison of the Original clause text vs. the Optimized clause text.
  • Actions — Approve & Apply (promotes the mutated clause directly to your active SOP) or Reject (dismisses the proposal). Only Owners and Admins can approve or reject.

  • SOPs — Manage and configure your workspace rules
  • Core Concepts — Understand the governance model and PCAS actions

The circuit breaker for AI agents