logo
Sign UpSign InTry for free
Agentic Workflow

Agentic workflows: BPMN for AI agent design

Agentic workflows are AI systems that plan, take actions, reflect on outcomes, and loop until a goal is met. As agents leave demos and enter regulated business processes, designing the workflow as a real diagram - with explicit branches, human-in-the-loop steps, and audit trace - becomes the difference between a working agent and a compliance liability.

Jack Finnegan, Founder & CEO, BA Copilot

By Jack Finnegan ยท Updated 21 May 2026

Map an agent workflowAI governance hub
What it is

What an agentic workflow actually is

An agentic workflow is an AI-powered process where one or more LLM-driven agents plan steps, call tools (functions, APIs, other agents), reflect on the results, and loop until they reach a goal. The pattern was articulated by Andrew Ng in 2024 in his 'agentic design patterns' work (reflection, tool use, planning, multi-agent collaboration) and has since become a dominant shape for AI systems beyond single-shot prompts.
Where traditional BPM defines the process explicitly and the human follows it, agentic workflows give the agent latitude to plan its own steps. That latitude makes them powerful for open-ended tasks but introduces real governance challenges: what tools is the agent allowed to call, when does a human review the output, how is the trace audited. Designing the workflow as a BPMN diagram is what makes those questions concrete.
The problem today

Agentic workflows live in code, not in diagrams - which is fine until they fail in production

Most agentic workflows are defined in code - a LangChain pipeline, a LlamaIndex agent, a custom orchestration framework. The behaviour emerges from the prompts and the tool list rather than from a designed diagram. This is fine in prototypes; it becomes a liability in regulated workflows where auditors need to see the decision flow, compliance needs to know when a human reviews, and the operations team needs to understand what happens when the agent fails.
The fix is treating the agentic workflow as a designed process. The BPMN diagram captures the planning step, the tool-call loop, the reflection task, the human-in-the-loop gateway, and the audit-trace task. The code implements the diagram; the diagram becomes the artefact that engineering, compliance, and operations can all read.
Four pillars

Four pillars of a working agentic workflow

Planning step

The agent decomposes the goal into expected steps before acting. Planning is what distinguishes agents from chatbots.

Tool use with explicit allowlist

The agent picks tools from an allowed set. The allowlist is part of the governance design - which tools, with which arguments, against which data.

Reflection and loop

After each tool call, the agent reflects: did this move toward the goal? The reflection step is what catches the dead-ends and prevents runaway loops.

Human-in-the-loop and audit trace

For high-stakes outputs, a human reviews before the response goes out. For all outputs, the trace is logged so auditors and operations can reconstruct decisions.

Process Map

An agentic workflow as a process map

The canonical loop - plan, pick tool, execute, reflect, loop until done, aggregate, human review, respond.

Open in editor

An agentic workflow as a process map

A canonical agentic AI workflow rendered as a BPMN 2.0 process. User goal, agent planning, tool selection, action loop with reflection, aggregation, and human-in-the-loop handoff.

  1. A user (or an upstream system) sends a goal to the agent.
  2. The agent plans - breaks the goal into a sequence of steps it expects to take.
  3. For each step, the agent picks a tool (function call) and executes the action.
  4. The agent reflects on the result against the original goal.
  5. If more steps are needed, the agent continues the loop. Otherwise it aggregates the results.
  6. For high-stakes outputs, the agent routes to human review before responding to the user.
  7. The final response is returned to the user; the trace is logged for audit and improvement.
What this diagram shows: The workflow starts when a user goal arrives. The agent plans a sequence of steps; for each step it picks a tool and executes the action; the reflect task evaluates the result against the goal. The more-steps gateway routes back into the pick-tool task for the next iteration, or forward to aggregation once the goal is met. High-stakes outputs flow through human review before responding. The trace is logged at the response step for audit and improvement.
FAQ

Frequently asked questions

What is an agentic workflow?

An agentic workflow is an AI system that plans steps, takes actions via tool calls, reflects on outcomes, and loops until a goal is reached. The pattern was articulated by Andrew Ng in his 2024 agentic design patterns work.

What's the difference between an agentic workflow and an AI chatbot?

A chatbot responds to a single prompt with a single response. An agent plans multiple steps, takes actions (via tool calls, function calls, or sub-agents), reflects on the results, and loops. The latitude to plan its own steps is what distinguishes agents from chatbots.

What is agent orchestration?

Agent orchestration is the layer that coordinates the planning, tool selection, reflection, and looping of one or more agents. Frameworks include LangChain, LangGraph, LlamaIndex, CrewAI, and Microsoft Agent Framework (Microsoft's convergence of AutoGen and Semantic Kernel, 1.0 GA April 2026), alongside increasingly first-party SDKs from the model vendors (Anthropic Claude Agent SDK, OpenAI Agents SDK, etc.).

Why design agentic workflows as BPMN diagrams?

Diagrams make the workflow auditable and reviewable. Code-only workflows hide the planning logic, tool allowlist, human-review checkpoints, and audit-trace integration inside imperative code. Diagrams expose them as artefacts auditors, compliance, and operations can read - which is increasingly required by the EU AI Act, NIST AI RMF, and ISO 42001.

How does this relate to AI governance?

Agentic workflows are a major source of operational risk in firms running AI at scale. AI governance frameworks (NIST AI RMF, ISO 42001, EU AI Act) require workflow-level documentation, control points, and audit traces. Treating the workflow as a designed BPMN process is the cheapest way to satisfy that requirement.

Does BA Copilot orchestrate agents?

No. BA Copilot is the modelling layer - it produces the BPMN diagrams that document the agentic workflow design. Orchestration frameworks (LangChain, LangGraph, etc.) execute the workflow. BA Copilot produces the artefact that ties the running code to the documented design.

Jack Finnegan, Founder & CEO, BA Copilot
From the founder

14 Years in BPMN

I'm Jack Finnegan. I've spent fourteen years working hands-on with BPMN, as an analyst, an engineer, and a product director, where I felt every sharp edge of legacy business process platforms.

BA Copilot is the platform I wanted on every one of these projects: AI-first process management, which treats BPMN as a first-class output rather than an export afterthought.

Jack Finnegan on LinkedIn
Cosmic background pattern
Decorative rectangle pattern

Make the agent design auditable

Open the agentic workflow template, model your agent design as BPMN - planning, tool calls, reflection, human review, audit trace - and produce the diagram engineering, compliance, and audit can all read.