Skip to main content

PRODUCT UPDATES

Release notes

The public changelog for London System Agent. This is a beta, so behavior keeps evolving, and only shipped changes are recorded here. The newest entry comes first.

Models · Execution

Gemini 3 support and sturdier Real runs

  • Added the latest Gemini 3 models (3.5 Flash, 3.1 Pro) and made Gemini 3.5 Flash the default. The model picker only lists models that actually run.
  • Transform agents (translate, summarize, classify…) now always act on your input — fixed a case where some models processed the instruction instead of the input.
  • Call hosted execution (POST /v1/runs) in Real mode with an issued API key; cost is debited safely from the key owner's credits.
  • Closed a race where concurrent runs could overdraw the balance, using an atomic budget reservation.
  • Switched refine/simulate to an async queue + polling, so a slow Real run that times out is never billed for work you didn't receive (no orphan billing).
Studio

A smoother execution flow

  • Graph compilation now streams over SSE — nodes appear on the canvas one at a time, showing compile progress in real time.
  • Moved the generation-mode toggle into Settings to simplify the main view.
  • Improved the studio dock layout.
  • Added an account dropdown menu.
Billing

Real-mode free tier

  • Signed-in users get a small one-time signup credit, so they can try a Real-mode run without paying.
  • Default execution mode is now Real on sign-in, with the execution mode surfaced front and center.
  • Fixed a runtime boot crash caused by a missing billing dependency.
Eval & gates

Pre-deploy evaluation gate wired up

  • Wired the policy editor, gate results, and the “Deployment blocked” indicator into the frontend (a visualization).
  • Added an NL Policy compiler that compiles natural-language policy into gate assertions.
  • Real-mode metered gate — per-run billing and an automatic stop when credits are exhausted.
  • Opened SQL storage paths for policy and gate results (SqlPolicyStore, SqlGateStore) so they persist in deployed environments.
  • Added node drag-to-reposition and a canvas overlay of evaluation results.
Eval

Deterministic eval engine and golden sets

  • Introduced a deterministic Release Gate engine with an N-run variance band and allowlist assertions.
  • Added golden datasets and a generic evaluator for the document, branch, and simple templates beyond research.
  • Labeled Stub reports as “contract regression” to avoid confusing structure checks with quality guarantees.
Runtime

Operational hardening

  • Added LLM call timeouts and endpoint rate limiting.
  • Masked PII in logs and introduced structured JSON logs correlated by run_id.
  • Added a tool-risk registry and a compile-boundary policy gate.
  • Added a graph node cap (20) and an LLM-suggested alternative when a refinement fails.
  • Persisted node layout, and introduced region-based automatic locale plus language change in Settings.
Content

Insights (blog) launch

  • Opened the public Insights section with notes on operable agent design — agent platforms, evaluation, and cost control.
  • Added screenshots of a real compiled graph and the Eval panel to the articles.
  • Added legal pages including the privacy policy and terms of service.