Skip to content

v3.5.0 — Bring your own data

Latest
Latest

Choose a tag to compare

View all tags
@MarkusNeusinger MarkusNeusinger released this 16 Oct 20:24
· 209 commits to main since this release
v3.5.0
7fcf2ff
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The window since v3.4.0 holds 57 merged pull requests (#12148 to #12209). Together with the 42 agent network PRs that v3.4.0 left for this release (#12099 to #12146), the release covers 99 PRs. The release is the "Use with my data" agent chat, which only admins can use for now. The full record is CHANGELOG.md, and each PR carries its own reasoning.

Added

  • Content-free usage telemetry for the agent chat. Four events, one wide JSON line each (#12205).
  • Agent usage card on /debug. Sessions, plots shipped, kept and repair rates over 7 or 30 days (#12205).
  • Excel, OpenDocument, Parquet, Feather, Arrow and NumPy files go into the agent chat's data box. At most 5 MB (#12194).
  • Data from your own code in "Use with my data". Python, JavaScript, R or a textbook formula, run in the sandbox (#12172).
  • The data field reads JSON-stat datasets from statistics offices. (#12173)
  • Agent results scored with the catalogue rubric. Restricted to the 16 criteria a user sees (#12157).
  • Plot text in many scripts. Noto fallback fonts in the renderer image: 37 files, 46.8 MB (#12144).
  • Agent cases: the triage half of quick feedback. A new agent_feedback table (#12142).
  • CI verifies every migration before merge. On a PostgreSQL 18 service container (#12142).
  • Quick feedback on an agent chat result stores the case for tuning. (#12139)
  • The assistant computes your data from a formula in step 1. (#12134)
  • The assistant changes your data in step 1, and you can undo it. (#12133)
  • The data field reads free text as a table, and you confirm it first. (#12130)
  • The "Use with my data" chat page, for admins. /debug/agent in front of the agent chat BFF (#12114).
  • A model-regression harness for the agent network. Runs eval cases through the real /v1 flow and exits non-zero when the pass rate drops (#12116).
  • The anyplot-renderer service that runs adapted plot code in Cloud Run sandboxes. Every theme of a job runs in a sandbox of its own, one at a time (#12115).
  • A run queue in front of every agent chat turn. (#12112)
  • The ADK runtime core of the agent network ("Use with my data"). The root agent, one adapter per library, a reviewer, and a bounded pipeline that always ends in a PlotResult (#12111).
  • Two-profile AST validator for the plot pipeline. The SECURITY and ADAPTATION profiles of the design (#12107).
  • Vertex AI budget guard. Disables aiplatform.googleapis.com once the month's cost reaches the budget amount (#12109).
  • Deterministic data handling for the agent network: parser, data roles, default bindings and the dataset store. (#12106)
  • Agent network design document. The planned "Use with my data" agent network on Google ADK 2.11 (#12099).

Changed

  • The privacy policy covers the data assistant preview. (#12208)
  • Three plot-quality rules from the agent chat's live tests, in the shared style guide and rubric. (#12203)
  • The plot of the day features standout plots and rotates without repeats. A featured pool of score 92 or more, no -basic spec (#12185).
  • The agent's adapter works to the reviewer's rubric, and fewer plots go to a repair. 95 % of the technical points skip the repair round (#12178).
  • Pasted data takes up to 1 MB, computed data up to 2 MB. (#12176)
  • The agent's plot pipeline reviews with the catalogue rubric and proofreads itself. Every round is reviewed and the best-scored version ships (#12165).
  • "Use with my data" runs in two steps: the data first, then the plot. (#12131)
  • The agent network's two services are deployed. anyplot-renderer and anyplot-agents run in Cloud Run (europe-west4) (#12127).
  • The fixed refusals speak eight languages. (#12126)
  • The agent pipeline reviews a repaired render once more. (#12122)

Removed

  • The unused agents-eval extra. Removes 29 packages from uv.lock and two Dependabot alerts (#12146).

Fixed

  • Pasted data keeps every digit, or a warning says what changed. (#12196)
  • A transient network glitch no longer leaves a page empty. (#12174)

Security

  • The agent chat's code validator closes exec and process routes through allowed modules. (#12177)
  • The render harness arms an audit hook in plot mode too. (#12177)
  • The agent validator closes module re-exports and private helpers. (#12124)

Full Changelog: v3.4.0...v3.5.0

Assets 2
Loading