AI Agents

Governed Quant Research Agent

Zeto extends beyond a research platform into a governed quant research agent for systematic experimentation. Local LLMs operate through MCP tools to inspect workflow state, build diagnostic context, retrieve research memory, review failure modes, propose controlled improvements, generate draft configurations, and interpret post-run results. The agent can reason across the research lifecycle, but it cannot bypass the engine: validation, approval, YAML rendering, and execution remain explicit, auditable, and human-controlled.

Natural-language research requests are transformed into structured research actions linked to deterministic diagnostics and provenance-aware artefacts. AI proposes; Zeto validates; the researcher approves; execution only occurs after explicit RUN authorisation.

Workflow Architecture

Five architecture layers — from research intent to governed execution to evidence-grounded AI review.

AI Planningintent → draft
Research Request

Natural language or structured request from researcher

Intent Parsing

Parse and classify intent into typed workflow actions

Research Planning

Select tools, data, configurations, and review strategy

Config Draft

Generate candidate experiment configuration draft

Human Governanceapproval gate
Human Approval

Researcher reviews draft, compares, and validates intent

Approved YAML

Final approved configuration is persisted

Human-Controlled Execution Boundary

Explicit human control before any execution is initiated. The LLM cannot run experiments or bypass approval.

Quant Research Enginesource of truth
Research Engine

Quant research platform executes the approved configuration

Research Artefacts

Metrics, diagnostics, plots, configs, and metadata are generated

AI Context Engineeringartefacts → context
Structured Context Engineering

Relevant artefacts, diagnostics, metadata, plots, and lineage assembled into structured context

Failure Mode Detection

Deterministically detect and flag issues before LLM review

AI Reasoning & Memoryreview → lineage
LLM Review

Review, comparison, and synthesis using structured evidence

Research Evolution Chain

Reviews, decisions, and insights linked into lineage and stored for future use

Execution is intentionally human-controlled. The orchestration layer assists research decisions but cannot execute experiments, modify results, or bypass approval.
Demo

End-to-end walkthrough from research request to evidence-grounded review and research evolution.

Governed AI Research Cycle DemoLocal Qwen operating Zeto through MCP tools: context building, failure-mode review, proposal, draft validation, approval, YAML rendering, RUN-gated execution, and post-run review.
Engineering Capabilities

Intent Routing

Natural-language research requests are converted into typed workflow actions before planning and review occur.

Context Engineering

Research artefacts are transformed into structured context containing diagnostics, metadata, plots, and failure modes before any LLM interaction.

Governed AI Review

Deterministic evidence is assembled before interpretation, ensuring reviews remain traceable, reproducible, and evidence-grounded.

Research Memory

Reviews, proposals, lineage records, and evolution chains preserve research state across iterations.

Governance
Governance Controls
  • LLM output is advisory
  • Human approval required
  • No autonomous experiment execution
  • Quant engine remains source of truth
  • Failure modes detected deterministically
  • Reviews linked to persisted artefacts