Alta3 Research Logo

Instructor-led training / ai-codex-mas-0H3

Enterprise Multi-Agent Systems in Codex

Build and validate an enterprise-pattern, multi-agent Codex workflow that produces evidence-backed, reviewable recommendations.
Request private trainingTalk to us
Instructor-led

ai-codex-mas-0H3

Duration
3 Hours
Format
Live Online
Access
Multiple Timezones

Overview

About this course.

This intensive, hands-on workshop teaches participants to engineer a working, enterprise-pattern multi-agent research system directly in Codex. Using a realistic consulting scenario, teams build a workflow that gathers evidence, validates calculations, challenges assumptions, and produces an executive-ready recommendation supported by traceable evidence.

This workshop covers the complete system lifecycle: workflow auditing, operating contracts, specialist subagent delegation, controlled tool integration, deterministic validation, adversarial review, and reviewable delivery.

Working in approximately 17 teams of three to four from a prepared but intentionally incomplete repository, participants complete and test a Codex-native Research OS workflow. The primary workflow delegates specialized work to Source Research, Benchmark Analysis, Calculation Check, and Red Team subagents defined in .codex/agents/. Agents exchange structured JSON artifacts for sources, claims, calculations, caveats, and open questions instead of relying on unstructured conversational handoffs.

Participants leave with a functional Codex workspace, reusable project assets, a packaged Codex Skill, and a validation pipeline that can be extended to enterprise consulting engagements.

Objectives

What you'll learn

  • Translate an audited workflow into an operating contract, repository instructions, a reusable Codex Skill, and machine-verifiable JSON schemas.
  • Configure native Python scripts and adapters for approved web retrieval, document access, and deterministic calculation verification in Codex.
  • Configure project instructions and delegation boundaries for specialist subagents operating in separate execution contexts with explicit return contracts.
  • Build validation gates that enforce schema compliance, citation metadata resolution, mathematical accuracy, and policy rules.
  • Generate an evidence-backed deliverable containing a decision, categorical confidence assessment, verified calculations, explicit caveats, and items requiring human review.
  • Separate deterministic code validation from probabilistic AI review.
  • Route unresolved findings, unsupported claims, policy violations, and conflicting or low-confidence recommendations to human review.
  • Test and repair multi-agent workflows using adversarial data fixtures.

Audience

Who it's for

  • Workflow Architects
  • Agent Engineers
  • Tool Engineers
  • Quality Engineers

Prerequisites

Before you start

  • Each participant must have their own Codex license.

Outline

Course outline.

Download PDF

1. Architecture Briefing and Baseline Run

  • Instructor-led walkthrough of the Research OS reference architecture.
  • Team role assignments for Workflow Architect, Agent Engineer, Tool Engineer, and Quality Engineer.
  • Hands-on exercise: execute the initial baseline test and observe failing validation assertions.

2. Operating Contracts and Schemas

  • Complete contracts/research-os.md to define system boundaries and operating rules.
  • Complete schemas/delivery.schema.json to enforce machine-readable output contracts.
  • Author skills/research-os/SKILL.md to package the reusable methodology.
  • Update the repository's AGENTS.md instructions.
  • Define structured artifacts for sources, claims, calculations, caveats, and open questions.

3. Specialist Subagents and Controlled Tools

  • Configure Source Research, Benchmark Analysis, Calculation Check, and Red Team specialist subagents in .codex/agents/*.toml.
  • Define delegation boundaries and explicit return contracts for separate execution contexts.
  • Extend tools/calculations.py with custom Python verification logic.
  • Configure approved retrieval and document-access adapters.
  • Hands-on exercise: test specialist subagents independently against sample data fixtures.

4. Primary Workflow Delegation and Structured Handoffs

  • Design the primary workflow orchestration.
  • Configure project instructions so the primary Codex workflow delegates specialized work to subagents.
  • Implement the structured handoff sequence: Input Payload → Specialist Subagents → Merged Draft.
  • Enforce strict return contracts so each specialist returns only the required results.
  • Aggregate specialist findings into a unified, reviewable output.

5. Deterministic Gates and Red-Team Review

  • Use schemas/delivery.schema.json to validate required fields, data types, and JSON structure.
  • Use tools/calculations.py to programmatically recompute numeric claims.
  • Extend validators/validate-delivery.py to enforce schema validation, citation resolution, and mathematical recomputation.
  • Verify that externally grounded claims resolve to valid source records with required citation metadata.
  • Configure .codex/agents/red-team.toml to perform adversarial semantic review.
  • Probe for unsupported assertions, missing caveats, and policy violations.
  • Route unresolved findings to a human reviewer.

6. Adversarial Fixture Testing and Repair

  • Run the workflow against conflicting/ data fixtures.
  • Run the workflow against missing-sources/ data fixtures.
  • Test failure handling for conflicting evidence and unresolved citations.
  • Hands-on exercise: iterate on agent instructions and validator assertions until failure paths execute cleanly.

7. Enterprise Procurement Workflow Lab

  • Build an approval API that accepts enterprise purchase requests.
  • Configure MCP connectors to enterprise contract repositories.
  • Configure MCP connectors to procurement policies and approved-vendor data.
  • Create Contract Analysis, Policy Compliance, Alternative Recommendation, and Red Team Review agents.
  • Route requests to the appropriate specialist subagents.
  • Evaluate purchase requests against enterprise contracts and procurement policies.
  • Recommend compliant suppliers or alternatives when appropriate.
  • Aggregate specialist findings into a single procurement decision.
  • Route low-confidence or conflicting recommendations to Human-in-the-Loop review.
  • Produce a structured, reviewable approval package.

8. Reviewable Delivery Pack

  • Generate output/delivery.json as the final Reviewable Delivery Pack.
  • Include an approval recommendation and confidence assessment.
  • Include contract and policy compliance statuses.
  • Include a recommended supplier or alternative.
  • Include supporting evidence and citations.
  • Include Red Team findings and deterministic validation results.
  • Flag whether human review is required.

9. Evidence-Pack Review and Close

  • Instructor-led inspection of generated delivery artifacts across teams.
  • Review the completed Codex workspace, reusable project assets, packaged Skill, and validation pipeline.
  • Discuss transferring the completed Research OS Skill to future client engagements.

For teams

Training a whole team?

Request private training for Enterprise Multi-Agent Systems in Codex and we'll follow up with a quote for your group.

Request a quoteBrowse the catalog