How an agentic workflow is built
A standard LLM request is one cycle: the user asks a question, the model generates an answer. An
agentic workflow breaks that linearity: the agent receives a goal — a task, not a question — and
builds its own plan for reaching it, calling tools, checking interim results and changing approach
where needed.
User goal
↓
Planning → list of subtasks
↓
Execute a step → function calling / MCP
↓
Reflection → goal reached?
├── No → adjust the plan → next step
└── Yes → answer the user
The four key patterns
Planning: the agent breaks a high-level goal into concrete steps before it starts executing. The
quality of that plan is the single biggest factor in how reliable the workflow turns out.
Tool use (function calling): the agent calls external APIs, databases and services through
function calling or MCP. That gives it access to current data beyond the model’s own knowledge.
Reflection: after each step the agent evaluates the result and decides whether a correction is
needed. Reflection is what stops errors from compounding.
Multi-agent delegation: complex tasks are handed to specialised agents over A2A, with an
orchestrator tracking overall progress.
An e-commerce workflow example
The task: “find me running shoes under $100 that can be delivered by Friday”.
- Planner: search → filter by price → filter by delivery → compare → select → order
- Search agent (MCP plus catalogue): 47 candidates
- Filter agent: 12 candidates under $100 with delivery by Friday
- Recommender agent: top 3 by rating and reviews
- Human-in-the-loop: shows the top 3, waits for the user to choose
- Checkout agent (ACP/YCP): places the order through a payment token
Risks and limits
| Risk | How to mitigate it |
|---|---|
| Hallucinations at intermediate steps | Verification against external sources, explicit checks |
| Infinite loops | A cap on the number of iterations |
| Irreversible actions | Human-in-the-loop before critical steps |
| Context leaking between agents | Minimise the data passed along |
Tip: the more complex the workflow, the more important it is to log every step. That is the only
way to debug a scenario in which the agent made the wrong call at step three out of seven.