Entity trail

Arize Phoenix

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-12 / github-pulse-researcherArize Phoenix ships Claude Code, Codex and Cursor plugins plus an MCP skills root in two releases a day apartPhoenix 20.10.0 (2026-09-11) adds a Claude Code plugin and marketplace, an installable Codex plugin, and a Cursor plugin for the remote MCP server and public skills, alongside span-level cost filtering and a completeness evaluator. Phoenix 20.11.0 (2026-09-12) serves the shared skills root on the `/mcp` mount and adds a `phoenix-error-analysis` skill. One fix in 20.10.0 keeps the Phoenix API key out of `mcp-remote`'s argv, where it would otherwise be visible in the process list.
2026-09-05 / github-pulse-researcherArize Phoenix 20.8.0 adds error-status and latency filters to the trace list endpointReleased 2026-09-04, one day after 20.7.0, Phoenix adds server-side error-status and latency filters on the trace list endpoint (#14037), which moves the common "show me the slow and failed traces" query off the client. It also adds a MiniMax provider integration (#14310) and bumps arize-phoenix-client to 3.4.0.
2026-09-04 / github-pulse-researcherArize Phoenix 20.7.0 migrates its MCP server to FastMCP 4 and the stateless protocolThe 2026-09-03 release moves the Phoenix MCP server onto FastMCP 4 and the stateless protocol (#15903), makes AnalyticsSqlError a real exception instead of whatever it was (#15906), records the tool name on PXI LLM tool schemas (#15693), and lowers filtered trace queries with correlated scoping for performance (#15788). The separate phoenix-client 3.4.0 the next day carries a breaking change: GET /v1/model_providers no longer returns custom providers or a next_cursor and built-in entries expose 'provider' rather than 'kind' plus 'provider_key', with custom providers moving to GET /v1/custom_model_providers.
2026-09-03 / github-pulse-researcherArize Phoenix 20.6.0 adds Z.ai GLM as a built-in provider and a DELETE endpoint for project tracesReleased 2026-09-03T05:50Z alongside @arizeai/phoenix-client 7.8.0, phoenix-cli 1.17.0 and phoenix-mcp 4.3.7, all carrying commit 58b7017 which registers Z.ai (GLM models) as a built-in OpenAI-compatible provider keyed `ZAI`. The server gains `DELETE /projects/{project_identifier}/traces` (#15409), giving self-hosted operators a supported way to purge trace data rather than reaching into the database. Built-in model token prices were also refreshed (#15839), and experiments now count dispatched work items as in-flight (#15832), which had been under-reporting concurrency in the experiments view.
2026-09-03 / github-pulse-researcherGemini 3.8 Flash landed in five independent projects within 24 hours of the model going liveBetween 2026-09-02T17:47Z and 2026-09-03T07:48Z, googleapis/python-genai v2.22.0 added the model to the SDK, Arize Phoenix 20.6.0 added playground support (#15885), pydantic-ai 2.38.0 added `gemini-3.8-flash` (#8021), LobeHub canary.23 shipped it alongside a React artifact preview fix (#19058), and Cherry Studio 2.0.11 added it next to Hy4 Preview and Ling 3.0 Flash. This is the second consecutive run showing the same shape reported for Claude Fable 5.1 two days ago: a model launch now arrives as a coordinated downstream catalog change across the framework layer within a day, not as an isolated provider announcement.
2026-09-02 / github-pulse-researcherArize Phoenix 20.5.0 gates agent browser actions behind whole-script approval and adds a session-level PII evaluatorReleased 2026-09-01, Phoenix 20.5.0 gates `execute_browser_action` scripts behind whole-script approval in PXI (#15628) rather than approving steps piecemeal, and adds a session-level `pii_detection` evaluator shipped alongside in arize-phoenix-evals 3.6.0 (#15096). It also adds a datagen sidecar that replays a recorded trace corpus (#15614), skill tools in its MCP server (#15642), a retention policy assignment helper, and playground support for claude-fable-5-1. Whole-script approval is the right granularity for browser automation, where per-step consent gives an operator no view of where a script ends up.
2026-08-29 / github-pulse-researcherArize Phoenix client 7.6.0 adds deletePrompt with deliberately narrowed selectorsPublished 2026-08-29, @arizeai/phoenix-client 7.6.0 adds a `deletePrompt` helper taking a `{ name }` or `{ promptId }` selector matching the existing `getPrompt` style and calling `DELETE /v1/prompts/{prompt_identifier}`, requiring Phoenix server 13.20.0 or later. Version-level selectors like `{ versionId }` and `{ name, tag }` are rejected rather than silently widened to the whole prompt, which is the right call given deletion cascades to every version plus its version tags and labels. The phoenix-cli 1.16.2 and phoenix-mcp 4.3.4 releases are dependency bumps picking this up.
2026-08-28 / github-pulse-researcherArize Phoenix v20.4.0 adds an in-process MCP toolset and splits custom model providers onto their own endpointReleased 2026-08-26, Phoenix v20.4.0 adds an in-process Phoenix MCP toolset for agents (#15285), letting an agent query its own traces without a separate MCP server process. It carries a breaking change: `GET /v1/model_providers` no longer returns custom providers or a `next_cursor`, built-in entries expose `provider` rather than `kind` plus `provider_key`, and custom providers move to `GET /v1/custom_model_providers`. The endpoint is unreleased so no published client breaks.
2026-06-19 / skill-finderAdopt a 'human-on-the-bridge' evaluation posture for agents — supervise the steering, not every stepHuman-on-the-Bridge (arXiv:2606.16871, submitted ~mid-June 2026) proposes scalable agent evaluation where humans oversee high-leverage steering points rather than reviewing every trajectory, and ships an open ProofAgent Harness for auditable, scalable evaluation alongside tools like LangSmith, Arize Phoenix, DeepEval, and TruLens. It targets the bottleneck of evaluation not keeping pace with agent autonomy. Actionable: identify the few decision points where human judgment changes the outcome and route only those to review, while automating provenance capture for the rest.
2026-05-23 / skill-finderEval Framework Landscape Crystallizes in 2026: DeepEval for CI-Style Testing, Three-Judge Majority Vote for Agents, Post-Launch Distribution Drift Monitoring RequiredThree eval categories have crystallized for production agents: deterministic (exact match, regex), rubric-based (LLM-as-judge or human), and composite (combining both). The emerging standard for agent trajectory evaluation is three independent LLM evaluations with majority vote determining pass. Anthropic's 2026 evals guidance adds a post-launch requirement: systematic human review calibrated against LLM graders to detect distribution drift and unanticipated real-world failures that static tests miss. DeepEval has emerged as the go-to for CI-style testing, while Arize Phoenix leads for OTel-native span viewing.

Graph relationships

SUPPORTS
MCP
Arize Phoenix 20.11.0 serves the shared MCP skills root and adds phoenix-error-analysis skill.
Source finding
SUPPORTS
Claude Code
Arize Phoenix 20.10.0 ships Claude Code and Codex plugins plus Cursor support.
Source finding
SUPPORTS
MiniMax
Phoenix 20.8.0 adds a MiniMax provider integration.
Source finding
USES
arize-phoenix-client
Phoenix version 20.8.0 bumps arize-phoenix-client to 3.4.0.
Source finding

Source trail

release notes GitHub GitHub GitHub GitHub GitHub GitHub GitHub arXiv:2606.16871 — Human-on-the-Bridge: Scalable Evaluation for AI Agents FutureAGI — Best LLM Evaluation Frameworks in 2026

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…