Agent Mode Guide

Agent Mode is NeuronFlow's durable workflow mode for longer-running automations that need memory, plans, checkpoints, and safe recovery between runs.

What Agent Mode changes

When Agent Mode is enabled, a workflow gains extra runtime behavior:

The workflow can keep a readable roadmap of what it is trying to do.

The system stores progress so the next run can continue from the right place.

Agents can read prior summaries, cursors, and context blocks from earlier runs.

Approval steps, waits, and retries behave more predictably in longer flows.

1. Plan support
2. Durable state
3. Memory and context
4. Checkpoint-friendly execution

Plan vs execution

The visible Plan is a roadmap, not the execution engine itself.

That means:

• the plan helps users understand what the workflow is trying to accomplish
• the executor still follows the node graph you built
• the plan is most useful for agent workflows that work in stages over time

Use the plan to explain the workflow, debug missing intent, and keep complex automations understandable for teammates.

Memory layers

Agent Mode works best when you understand the three main context layers:

1. Input

The payload coming from the previous node or trigger.

2. Memory

Durable values stored with the Memory node.

Good examples:

• pagination cursors
• sync checkpoints
• last processed record id
• agent state flags

3. Context Memory

Reusable knowledge blocks stored with the Context Memory node.

Good examples:

• brand voice
• approval policy
• customer-specific instructions
• SEO rules

Recommended build pattern

For most agent workflows, this pattern stays dependable:

1. Trigger
2. Context / Memory preload
3. AI Agent
4. Tool nodes
5. Switch / Approval / Wait
6. Memory update
7. Output

This keeps the agent grounded, prevents repeated work, and makes future runs easier to resume.

Best uses for Agent Mode

Agent Mode is strongest when the workflow:

• spans multiple steps
• needs human approval
• needs to remember something across runs
• may pause and continue later
• benefits from a readable roadmap

Great examples:

• inbox triage and follow-up
• lead enrichment over time
• SEO monitoring and reporting
• multi-step content drafting
• recurring operations reviews

When not to use Agent Mode

You usually do not need Agent Mode for:

• one-shot API fetches
• simple trigger → tool → output flows
• very small utility workflows
• workflows that do not need memory, planning, or resumability

If the job is short and deterministic, standard workflow mode is simpler and easier to maintain.

Common mistakes

Forgetting to store progress

If the workflow needs continuity, save a cursor or checkpoint with the Memory node.

Overloading the prompt

Do not keep stuffing everything into one AI prompt. Move stable instructions into Context Memory.

Treating plan as execution logic

The plan helps people understand the workflow. The node graph still controls the actual execution.

Publishing before testing recovery paths

Always test:

• approval pauses
• wait steps
• retry behavior
• the next run after an interrupted execution

Operator checklist

Before publishing an agent workflow:

1. confirm the workflow passes Preflight
2. confirm the AI model is selected
3. confirm tool connections are valid
4. confirm memory keys and sessions are intentional
5. confirm the workflow can recover cleanly after a pause or failure

Good practice

Build the first version in Development, validate in Staging, and only then move it to Production once the plan, state, and recovery behavior all make sense.