Agents, Tools & Skills
Turn retrieval into action. Turing's configurable agents call your tools (custom Groovy with HTTP, search and slot helpers), run Anthropic-standard skills in an isolated Docker sandbox, and federate across MCP servers — while Turing can itself be exposed as an MCP server so any MCP-aware client can query your content. Each provider's own server-side tools (web search, code execution, computer use, image generation) are available too, opt-in per agent.
For anything beyond a single turn, the declarative chat-flow DSL orchestrates multi-step conversations — slots, form capture, A/B experiments, human-in-the-loop approvals, webhooks and scheduled routines. An evaluation harness lets you gate agent and prompt changes in CI before they ship.
AI Agents
AI Agents are purpose-built specialists you compose from an LLM, tools, and a persona. They are the conversion engine of Turing ES.
Tool Calling
Native tool calling capabilities available to AI Agents in Turing ES — DSL-based search tools, RAG, web crawling, and more.
Capabilities
The capability registry and two-level gate in Viglet Turing ES — how provider-native server tools, request options, and Turing's own tools surface in one taxonomy and run in one tool loop.
Client Tools & Agent↔UI (AG-UI)
Live tool-call events, browser-executed client tools, the park/resume protocol, and generative UI in Viglet Turing ES — the agent↔UI surface.
Custom Tools
Build your own AI tools in Groovy and let your agents call them like first-class capabilities. Plus Vibe Coding — the AI helps you write the tool, describe it to the LLM, and shape its parameters.
MCP Servers
MCP works both ways in Turing ES — connect AI Agents to external tool servers (client), and expose Turing's own search/RAG/agent/analytics tools to Claude Desktop, Claude Code, or Cursor (server).
Skills
Run Anthropic-compatible Skills inside Turing ES. A Skill is a portable folder (SKILL.md + scripts + references) that runs in a hardened Docker sandbox, activated by progressive disclosure, and usable in both chat and semantic navigation.
Agent Workspace
A per-conversation blob store that lets an agent build, keep, and hand back files — drafts, reports, charts, offloaded tool results — without bloating the prompt.
Chat Flow
Author conversational state machines visually. A rich node catalog (questions, slots, tools, routines, webhooks, approvals, sub-flows), three guardrail strategies, error edges, a typed slot store, and AI-assisted authoring (Vibe Coding) that lets you describe the flow in plain English and watch it draw itself.
Human-in-the-Loop
Put a person in the conversation when it matters. Pause a flow for an explicit human approval over email / Slack / webhook, or watch a live conversation as a spectator and take the wheel mid-chat as a co-pilot.
Webhooks
Push chat conversations to your CRM or automation tools. Fire a webhook automatically when a slot is filled, explicitly as a handoff channel, or deterministically from a flow node — with payload templates, encrypted credentials, and HMAC signing.
Routines
Run long work asynchronously inside a conversation. A routine is a named async job fired by a chat-flow scheduleAgent node — it runs on a queue while the chat turn returns instantly, then writes its result into a slot to resume the flow.
Experiments (A/B Testing)
A/B test your chat flows natively. Split traffic across variants by weight, measure the winner with a real significance test, let a multi-armed bandit shift traffic automatically, and auto-promote the champion — all declared on the flow itself.
Agent Evaluation (CI)
Regression-test your AI agents like code. Capture golden conversations with expected slots and outcomes, replay them through the real engine, score them with a deterministic diff plus an LLM judge, and block a flow from publishing when it regresses.