Enterprise DNA
EDNA Open Playbook

AI Agents That Actually Work

The open Enterprise DNA playbook for designing, building, evaluating and deploying agents inside real businesses.

An agent is only useful when it can see the work, do the work and prove the work was done.

Start with the introduction View the labs
Read on GitHub PDF and EPUB at launch v1.0.0 Updated 23 July 2026 Free and open

Business-ready agent

  1. 1 Model
  2. 2 Context
  3. 3 Tools
  4. 4 Workflow
  5. 5 Verification

See the work. Do the work. Prove it was done.

chapters
11
chapters
labs
8
labs
verified
0
verified
paths
3
paths
to read
~5h
to read
free & open
v1.0.0
free & open

Choose your path

A path highlights the chapters that fit your job. It never hides the rest, and you can switch at any time.

Executive

Understand where agents pay off, what can go wrong and what a credible deployment requires.

6 chapters · about 2h

Operator

Turn a real process into a controlled agent workflow.

9 chapters · about 4h

Builder

Work through the full architecture and every core lab.

11 chapters · about 10h

The map

Foundation

Design

Control

Deployment

Scale

Chapters

  1. The Demo Is Not the System

    12 min · foundation

    Understand why agent reliability lives outside the model, and what the rest of this playbook will build.

    Executive Operator Builder

  2. What an AI Agent Actually Is

    22 min · foundation

    Build a practical mental model of an agent without the hype, so you can reason about what it will and will not do.

    Executive Operator Builder Lab: Map an Agent-Worthy Workflow

  3. Pick Work Worth Automating

    28 min · foundation

    Find the tasks with real value, clear boundaries and outcomes you can actually check.

    Executive Operator Builder Lab: Map an Agent-Worthy Workflow

  4. Give the Agent the Right Context

    35 min · intermediate

    Design what the agent sees, when it sees it and what should stay out.

    Operator Builder Lab: Build a Research and Brief Agent Lab: Build a Document-to-Report Agent

  5. Give It Tools That Are Safe and Clear

    32 min · intermediate

    Create tool contracts that let the agent act without creating chaos.

    Operator Builder Lab: Build a Document-to-Report Agent

  6. Design the Workflow and Control Loop

    35 min · intermediate

    Manage state, retries, delegation, stopping conditions and proof of completion.

    Operator Builder Lab: Build a Data-to-Decision Agent

  7. Add Memory Without Creating a Mess

    30 min · intermediate

    Use session memory, user memory and knowledge retrieval deliberately, not by default.

    Builder

  8. Build Guardrails and Human Approval

    30 min · intermediate

    Decide what the agent may do on its own, what needs a human, and how it escalates.

    Executive Operator Builder Lab: Add Human Approval to a High-Risk Action

  9. Evaluate Before You Trust It

    40 min · advanced

    Build tasks, expected outcomes, graders and regression checks so you know it works.

    Executive Builder Lab: Build an Agent Evaluation Harness

  10. Deploy, Observe and Improve

    35 min · advanced

    Ship the system, inspect traces, control cost and learn from real failures.

    Operator Builder Lab: Trace and Diagnose a Failed Run

  11. Scale Into an AI Operating Layer

    30 min · advanced

    Move from one working agent to a coordinated operating system for the business.

    Executive Operator Builder Lab: Capstone: Multi-Step Operations Agent

Labs

All labs →
Guided 45 min · beginner

Map an Agent-Worthy Workflow

Take a real process from your own work, score it against clear criteria, and decide whether it deserves an agent.

No codeSpreadsheet or worksheet
Guided 90 min · intermediate

Build a Research and Brief Agent

Build an agent that researches a topic from real sources, cites them, and only reports done when the brief passes a completion check.

PythonOne model provider
Guided 90 min · intermediate

Build a Document-to-Report Agent

Build an agent that reads a pile of messy source documents and writes a clean structured report, then proves every section maps back to a real source.

PythonOne model provider
Guided 90 min · intermediate

Build a Data-to-Decision Agent

Build an agent that reads a CSV, computes the numbers, draws a chart, and checks its own figures before it reports a decision.

PythonOne model provider
Guided 75 min · intermediate

Add Human Approval to a High-Risk Action

Wrap a high-risk action like sending an invoice or issuing a refund in a propose, hold, approve, act state machine, with an audit record for every step.

PythonOne model provider
Runnable 90 min · intermediate

Build an Agent Evaluation Harness

Build a repeatable test set that runs your agent over known tasks, grades the results three ways, and produces a pass or fail regression report.

Pythonpytest

What you will leave with

  • An agent opportunity map for your own work
  • A context design template
  • A tool contract checklist
  • A human approval and escalation map
  • An evaluation set template
  • A deployment readiness checklist

Free and open on GitHub

The full playbook, every lab and all templates live in a public repository. Read it, fork it, open an issue or suggest an edit. Contribution does not guarantee publication, but good work gets in.

Content: CC-BY-4.0 Code: Apache-2.0 Enterprise-DNA-OS/ai-agents-that-work
Read on GitHub

Learn it with structure

The playbook is the open reference. EDNA Learn gives you courses, practice and MENTOR through the full skill path.

Build with other practitioners

Built a lab? Show the work. Share the architecture, what broke and what you changed with other EDNA Builders.

Deploy it in your business

Have a workflow worth automating? We can map it, build the first agent and put the verification and human controls around it.