Skip to main content
Hyde Workshop Human-governed AI

AI Lab // Six-role evaluation bench

Choose the workflow before you choose the AI.

Review AI models, coding agents, workspace tools, and open standards by what they can read, change, and send, and by how much human oversight a real deployment needs. Every review below uses the same six A.M.A.N.D.A. checks and links to the vendor's own documentation.

Run the Workflow Diagnostic Read the AI Forge security controls
AI Aria, the Hyde Workshop audit role, in the laboratory uniform

How we review

Each system gets the same six checks, one per A.M.A.N.D.A. role: Mission, Evidence, Audit, Routing, Defense, and Decision. We link to primary sources and note when a review was last updated.

  • Oversight needed is our editorial judgment of how much human review a typical deployment should keep, based on what the system can read, change, or send. It is not a vendor rating or a benchmark score.
  • Model names and features change often. Confirm the current version on the vendor's page before you buy or build.

Filter by category

Systems under review

  • Coding agent and models

    Claude and Claude Code (Anthropic)

    Anthropic's Claude models, with Claude Code for agentic coding from the terminal, web, desktop, and mobile. Current public models are Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5.

    Best for
    Multi-step coding and long documents
    Can change
    Files and commands you allow
    Oversight needed
    High
    A.M.A.N.D.A. review
    A
    Amy: MissionSet the folders, branch, and deliverable before the agent starts.
    M
    Maya: EvidenceKeep the terminal log, generated tests, and dependency changes as the record of what ran.
    A
    Aria: AuditCheck that the build passes and the change fits the existing architecture.
    N
    Nancy: RoutingSend every change through your normal pull-request and review pipeline.
    D
    Dawn: DefenseRun in a container or VM, keep secrets and .env files out of reach, and restrict network access.
    A
    Astra: DecisionPilot in an isolated environment. A person approves every merge.

    Primary source: Claude documentation (opens in new tab)

  • Models and API

    Gemini and the Gemini API (Google)

    Google's Gemini models, available in the Gemini app and through the Gemini API for developers. Check Google's documentation for the current model names and limits before you build.

    Best for
    Large-context and multimodal tasks
    Can change
    Whatever your tools and tokens allow
    Oversight needed
    Medium
    A.M.A.N.D.A. review
    A
    Amy: MissionDefine the bounded task and exactly what "done" means before running long jobs.
    M
    Maya: EvidenceLog tool calls, inputs, and outputs so each result can be traced to its source.
    A
    Aria: AuditTest the same prompts repeatedly and compare outputs with a known-good sample.
    N
    Nancy: RoutingSet timeouts, retry limits, and a human escalation path for jobs that stall.
    D
    Dawn: DefenseStore API keys server-side, scope tokens narrowly, and validate structured output before it touches a database.
    A
    Astra: DecisionAdopt for analysis work first. Any action that changes data needs a human checkpoint.

    Primary source: Google AI for Developers (opens in new tab)

  • Workspace and API

    ChatGPT workspaces and the OpenAI API

    OpenAI's ChatGPT for individuals and teams, plus the OpenAI API for developers. Check OpenAI's documentation for current models, plans, and data controls.

    Best for
    Shared team knowledge work
    Can change
    Connected apps you approve
    Oversight needed
    Medium
    A.M.A.N.D.A. review
    A
    Amy: MissionWrite the workspace rules: who has access and which workflows the assistant may run.
    M
    Maya: EvidenceRead the plan's data-retention and training terms, and record which version you reviewed.
    A
    Aria: AuditCompare answer quality against speed on your own tasks, not on vendor demos.
    N
    Nancy: RoutingConnect only approved apps, with permissions set per person.
    D
    Dawn: DefenseBlock unapproved data export and review third-party connections regularly.
    A
    Astra: DecisionAdopt for internal knowledge tasks. Anything customer-facing gets an editor's review.

    Primary source: OpenAI documentation (opens in new tab)

  • Open standard

    Model Context Protocol (MCP)

    An open protocol for connecting AI applications to tools and data sources. Each MCP server you add gives the assistant new things it can read or do.

    Best for
    Connecting assistants to your own tools
    Can change
    Whatever each server exposes
    Oversight needed
    High
    A.M.A.N.D.A. review
    A
    Amy: MissionList the one task each server exists to support. Skip servers without a clear job.
    M
    Maya: EvidenceRecord who publishes each server, its source, and the version you installed.
    A
    Aria: AuditRead the tool list. Separate read-only tools from tools that write, send, or delete.
    N
    Nancy: RoutingMap which assistant can call which server, and who owns each connection.
    D
    Dawn: DefenseTreat tool output as untrusted: web pages and documents can contain instructions aimed at the assistant (prompt injection). Use minimum permissions.
    A
    Astra: DecisionApprove servers one at a time. Require a person to confirm any action that writes, sends, spends, or deletes.

    Primary source: modelcontextprotocol.io (opens in new tab)

  • CMS infrastructure

    WordPress 7.0 AI Client and Connectors

    WordPress 7.0, released 20 May 2026, adds an AI Client, a Settings > Connectors screen for provider keys, and the Abilities API. Core does not bundle any AI provider, so plugins call the providers you connect.

    Best for
    Standardizing AI features across plugins
    Can change
    Site content, depending on the plugin
    Oversight needed
    High
    A.M.A.N.D.A. review
    A
    Amy: MissionName the exact editorial or admin feature you need before enabling any AI plugin.
    M
    Maya: EvidenceCheck each plugin's declared WordPress 7.0 support and where it sends data.
    A
    Aria: AuditReview REST endpoints and user capabilities, and confirm injected UI causes no layout shift.
    N
    Nancy: RoutingMap the data path from the editor, through the Connector, to the external provider.
    D
    Dawn: DefenseKeep keys server-side. Do not let public forms or comments feed prompts unfiltered. Prefer fixed REST endpoints over arbitrary client prompts.
    A
    Astra: DecisionTest on staging first. Watch API cost and error logs before production.

    Primary source: WordPress Core development blog (opens in new tab)

Reviews last updated 16 October 2026. Suggest a system to review through the contact page.