Understanding AI Agents in MFO¶
AI agents add intelligence to workflows.¶
In the case of the use of LLM models, there are two types of calls:
1: Simple calls (requesting a summary, extracting entities, writing text, OCRing an image, transcribing an audio file, etc.)
2: Agent calls (ask to reflect, make a choice, choose a tool, call another agent, etc.).
The choice is implemented according to the needs of a workflow. AI agent calls are often more complex. Agents could work in team mode, or in working patterns explained in Anthropic's guide to effective agents.
These agents can:
- Interpret natural language.
- Make informed decisions in uncertain situations.
- Collaborate with deterministic tasks to enrich workflows.
AI Agent as a Dynamic Workflow¶
With MindFlight AI Orchestrator, you can think of agents as dynamic workflows (in teams) or dynamic task flows (solo). Instead of following a predefined sequence, the AI agent has the freedom to adapt within a certain framework.
Patterns of AI agents workflows that can be managed with MFO:¶
| Pattern | Explanation |
|---|---|
| Augmented LLM | Combines the capabilities of an LLM with external data or tools to enrich responses. |
| Prompt chaining workflow | Chains multiple prompts where each response serves as input for the next, creating a sequential flow. |
| Routing workflow | Directs queries to different agents or tools based on predefined conditions. |
| Parallelisation workflow | Executes multiple tasks simultaneously to speed up processing or compare outputs. |
| Orchestrator-workers workflow | An orchestrator agent coordinates multiple specialised agents to accomplish subtasks. |
| Evaluator-optimizer workflow | Evaluation and optimisation loop where one agent analyses and adjusts the results of another. |
| Autonomous agent | Agent capable of independently planning and executing a series of actions without supervision. |
Design patterns of AI agents¶
| Pattern | Explanation |
|---|---|
| ReAct Agent | Alternates between reflection and actions, allowing external tools to be used during execution. |
| CodeAct Agent | Generates, executes, and corrects code iteratively to achieve a goal. |
| Tool use | Explicitly selects and uses external tools (APIs, scripts, databases). |
| Self-Reflection | Analyses its own outputs to detect errors and suggest improvements. |
| Multi-agent workflow | Several specialised agents collaborate and communicate to accomplish a complex task. |
| Agentic RAG | Retrieval-Augmented Generation architecture where agents retrieve data and generate enriched content. |
One ring to rule them all¶
All these patterns can be implemented using workflows and the creation of specific tools in MFO providers.
This is one of the strengths of the MFO framework over nocode and low-code tools. The flexibility with which workflows can be defined means that an agent collaboration mode can be tested and iterated more quickly.
LLM providers¶
In MFO, AI is seamlessly integrated via any provider (e.g. OpenAI, Anthropic, Grok, Groq, Mistral, DeepSeek, ...) or local server. It is possible to integrate multiple providers in the same workflow. You can use the AI models that best suit your needs and transaction processing costs.
If you have stringent security requirements, you can run local models and connect them to MindFlight AI agents without any issues.
How AI Agents Use Internal Memory Tools in the MFO Framework¶
Imagine each AI agent in the MFO framework as a highly skilled team member who not only follows instructions but also remembers, learns, and improves over time.
🧠Memory: The Agent’s Short-Term Brain¶
Every agent needs a way to keep track of important details—just like a person taking notes during a meeting. The MFO Server gives agents a memory system (🔗 see "Memory system of the MFO Server") that acts like a personal notebook. For example:
- When solving a customer request, the agent writes down the current status.
- If asked later, “Where did we leave off?”, it quickly retrieves the exact point and continues smoothly.
This memory is scoped by user and task, keeping everything organized and private, ensuring agents don't mix up information between different users or projects.
🔍 RAG: The Agent’s Long-Term Library & Learning Hub¶
Beyond short-term memory, agents also need a deep library of knowledge to draw from. That’s where the RAG (Retrieval-Augmented Generation) system comes in (🔗 see "RAG and MFO Server")). Think of RAG as a smart filing cabinet:
- Whenever new documents are added (manuals, reports, FAQs), they are broken into pieces and deeply analyzed so agents can later find exactly what’s needed in seconds.
- When an agent receives a complex question, it searches this library, finds the best pieces of information, and uses them to craft precise answers.
💡 What’s exciting? This library isn’t static—it learns which pieces are the most useful. Each time an agent solves a problem successfully, the system takes note and reinforces the best sources. Over time, it becomes smarter and more reliable—like a librarian who remembers which books helped the most in the past.
🔄 Orchestration: The Agent’s Personal Assistant¶
Agents don’t just handle one thing—they often work on complex projects with many steps. The MFO tools allow agents to:
- Coordinate different tasks (e.g., analyze an email âž” store the result âž” tag the conversation).
- Collaborate with other agents, sharing results through memory or tapping into RAG when deeper knowledge is needed.
All of this happens seamlessly, with the agent acting like a project manager—organizing tasks, checking progress, and making sure everything runs on time.
Summary¶
The MFO framework turns AI agents into autonomous, evolving team members. They don’t just execute—they think, remember, learn, and optimize. This ensures your workflows are not only automated but also adaptive and improving continuously.
Why It Matters:¶
AI agents expand what's possible. They let workflows handle fuzzy tasks—things a traditional script would struggle with—making your processes smarter and more adaptable.