Back to Articles
Technical Guide March 7, 2026 3 min read

From Workflow to Agents

By Fredrik Brattén

AI Agents Large Language Models (LLMs) CrewAI LangGraph Multi-Agent Systems RAG (Retrieval-Augmented Generation) HTML
Cover image for: From Workflow to Agents

Resources

  • AI Agentic Workflows (DeepLearning.AI) Andrew Ng's foundational series on how iterative agentic patterns and multi-step workflows outperform zero-shot LLM prompts.
  • LangGraph: Multi-Agent Workflows Documentation for building stateful, multi-agent systems that utilize the 'contracts and handoffs' logic described in the article.
  • CrewAI: Orchestrating Role-Based Agents A framework specifically designed to define specialized roles (like Historian or Writer) and manage their interactions within a workflow.
  • Microsoft AutoGen An open-source framework for building multi-agent systems where agents can talk to each other to solve complex tasks.

Tech Stack

AI AgentsLarge Language Models (LLMs)CrewAILangGraphMulti-Agent SystemsRAG (Retrieval-Augmented Generation)HTML

Key Takeaways

  • Workflows become candidates for agentic delegation only when their internal steps are clearly defined and their inputs and outputs are predictable.
  • Deconstructing complex processes into specialized, modular agent roles improves output quality by focusing each agent on a single, well-defined phase of the workflow.
  • Implementing scoped access ensures that agents only receive the specific data required for their task, effectively preventing scope creep and unintended system actions.
  • Replacing open-ended conversations with structured handoff contracts between agents enhances traceability and allows individual modules to be validated or swapped independently.

Who this is for

AI developers and operations leads designing multi-agent systems for workflow automation

When Workflows Stabilize

A workflow becomes a candidate for delegation once it meets two criteria:

  • The steps are defined clearly enough that someone else could follow them
  • The inputs and outputs of each step are predictable

The newsletter process (Part 2) reached this point after a single use. Each phase - from source gathering through humility audit to final formatting - had clear boundaries, clear inputs, and verifiable outputs.

That makes it agentizable.


Five Agent Roles

The newsletter workflow splits naturally into five specialized roles:

Historian (Phase 1-2)

Finds origin material, prior conversations, old prompts, and project lineage. Gathers the raw context that everything else builds on.

Access: project docs, selected prior chats, repo metadata

Structurer (Phase 3)

Turns raw material into architecture. Defines section order, content mapping, and information hierarchy. Produces the blueprint that the Writer follows.

Access: Historian output, site structure, prior newsletter templates

Writer (Phase 4)

Generates the draft within the established structure and scope. Follows tone rules and formatting constraints. Does not decide what to include - that was already decided.

Access: Structurer output, tone guidelines, source content

Auditor (Phase 5-6)

Runs the humility audit. Checks for overpromising, absolute language, misleading claims, tone mismatches. This role is not optional - it is a required gate before output.

Access: Writer draft, scope document, humility checklist

Publisher (Phase 7)

Formats the audited draft for the target platform. Produces copy-paste HTML, social post variants, subject lines. Handles the last mile between "done" and "deployed."

Access: Auditor-approved draft, platform specs, brand assets


Scoped Access, Not Full Access

A critical design decision: agents should not have full access to everything.

Each role receives only what it needs:

  • The Historian can read project docs but cannot modify them
  • The Writer receives the structure but not the raw source - preventing scope creep
  • The Auditor sees the draft and the scope document - enough to verify alignment, not enough to rewrite
  • The Publisher receives only the approved final version

This is not bureaucracy. It is how you prevent agents from doing things they were not asked to do.


Contracts, Not Conversations

Agent handoff sequence with scoped access

Agent handoff sequence with scoped access

Each agent-to-agent handoff is defined as a contract:

From To Handoff
Historian Structurer Source inventory + context summary
Structurer Writer Section blueprint + content mapping
Writer Auditor Draft + scope document
Auditor Publisher Approved draft + audit notes

When agents communicate through structured contracts rather than open-ended conversations, several things improve:

  • Failures are traceable - you can see exactly where the chain broke
  • Roles are replaceable - swap one agent for another without redesigning the pipeline
  • Quality is measurable - each handoff can be validated against its contract

Beyond Newsletters

The same pattern applies to any repeatable creative or technical workflow:

  • Blog post production
  • Documentation generation
  • Report creation
  • Social media content derivation
  • Slide deck assembly

The agent roles may shift - a Code Reviewer replaces the Auditor, a Diagram Generator replaces the Publisher - but the structure remains.

Stable workflow. Clear roles. Scoped access. Verifiable handoffs.

That is how workflows become systems.


This article is Part 4 of the From Meta-Prompt to Asset Factory series on Adaptivearts.ai.

Previously: Building While Publishing: Turning Work Into Assets - every project should emit reusable materials. Next: SPINE as the Structural Layer - where 5PP lives inside SPINE.

Share this article

Tags

#ai agents#workflow automation#agentic workflows#content automation#ai orchestration