Skip to main content
Version: 8.8 (unreleased)

Design and architecture

Plan and design your agentic orchestration solutions, and understand recommended architecture guidelines.

Plan​

Follow these principles when planning your agentic orchestration solution:

  • Problem first: First, identify any problem you might have in a process, and only then determine whether an AI agent could help solve the problem. Do not use an AI agent where it is not really necessary, or just for the sake of it.

  • Architect for composability. Avoid becoming too dependant on a specific LLM model, for example by doing too much fine tuning. This allows you to more easily integrate newer LLM providers and models in the future that better suit your needs.

  • Observability and governance: Use Operate and Optimize for visibility into your agentic orchestration processes.

When to use deterministic or non-deterministic orchestration​

Agentic orchestration involves blending both deterministic and dynamic (AI-driven) process orchestration into your end-to-end processes. It is important to understand when to use each approach:

 DeterministicDynamic
Suitable for

  • Clear, predictable paths.

  • Repeatable, rules-based decisions.

  • Fast, efficient execution.

  • Regulation dictates execution sequence.

  • Flexible, context-aware choices.

  • AI-driven task planning.

  • Unstructured cases: classification, triage, or investigation.

  • Adaptable to data and conditions.

Enabled by

  • Advanced workflow patterns.

  • Business/IT alignment on process and decision logic.

  • Dynamic task scheduling and tracking.

  • Collaboration between AI and Human.

info

To learn more about determining when and where to use AI agents within your automation strategy, download the why agentic process orchestration belongs in your automation strategy guide.

Design and architecture​

Follow these principles when designing your agentic orchestration solution:

  • Guardrail sandwich: Apply guardrails in your process when using agents. For example, you could have one agent performing the task execution, with another agent following up to check the chain of thought and make sure every execution is compliant. If the execution is not compliant, route to a human for additional validation.

  • Human-in-the-Loop escalation: Provide an agent with an escalation path to a human - confidence levels are useful, but it is good to always provide deterministic outbreaks for agents.

  • Prompt versioning: Version every prompt, so you can revert to using a previous prompt when required.

Mixing Agents with Workflow patterns​

1Process Flow Within Tools: Full BPMN control inside ad-hoc subprocesses for ultimate flexibility.
2Agents pivot instantly with external messages and timers: During execution, agents can be influenced by events like external messages or timers, enabling on-the-fly adjustments.
3Event-Driven Agent Reconfiguration: Sub-workflows handle new data, guiding the next AI steps.
4Agents orchestrate sub-workflow: A tool doesn't need to be a single tool - it can be a whole subprocess.
5Multi-Agent Orchestration: Agents orchestrate other agents for streamlined, scalable solutions.