AI Agent Workflow Automation: How AI Agents Execute Multi-Step Tasks

Most “AI automation” is just an API call with a prompt. Send text in, get text out. That works for summarization and translation, but it falls apart the moment a task requires more than one step.

Real work looks like this: read a folder of files, decide which ones need processing, transform them, write the results somewhere, and handle errors along the way. That requires an agent — an LLM that can use tools, read the results, and decide what to do next.

NORA ships with four AI agent node types, each designed for a different complexity tier. All four run locally on a Windows desktop. No cloud infrastructure. No container orchestration. No API gateway.

What Makes an Agent Different from an API Call

A standard LLM API call is stateless. You send a prompt, you get a response. If you need multiple steps, you write the glue code yourself — parse the output, call the next API, handle failures, manage state.

An agent adds two things on top of the LLM:

Tool use — the LLM can request that specific tools be executed (read a file, run a command, call an API)
Decision loops — after each tool result comes back, the LLM evaluates it and decides the next action

This loop — think, act, observe, repeat — is what turns a language model into something that can complete multi-step tasks without human intervention at each step.

NORA implements this pattern across four node types.

Level 1: AI Agent Node — LLM with Tool Routing

The AI Agent Node connects an LLM to downstream workflow nodes. The LLM receives a prompt, decides which tool to invoke, and NORA routes execution to the corresponding downstream node.

Key capabilities:

Single-turn or multi-turn dialogue. Configure max conversation turns for iterative refinement.
Tool integration. Each downstream connection is a tool the LLM can invoke. The LLM decides which one based on the input.
Memory folders. Point the node at a directory of context files (.txt, .md, .json, .html). The LLM reads these as background knowledge before making decisions.
Tool parameters from upstream nodes. Pass dynamic values from earlier workflow steps into tool calls.
Conversation history persistence. Multi-turn sessions retain context across turns.
Email summaries with token cost. Get a Gmail notification with what the agent did and what it cost.

When to use it: Tasks where the LLM needs to choose between a small set of known actions — classify a document and route it, triage an incoming request, decide whether to approve or escalate.

Three AI providers available: OpenAI (GPT-5.5, GPT-5.4, mini, nano), Anthropic (Claude Opus 4.7, Sonnet 4.6, Haiku 4.5), and Google (Gemini 2.5 Flash/Flash-Lite/Pro plus Gemini 3 preview variants). Configure per node with your own API keys.

Level 2: AI Autonomous Agent — Full Autonomy with Safety Limits

The AI Autonomous Agent Node is a self-directing agent. Give it a goal, and it figures out the steps. The LLM decides which tools to call, executes them, reads the results, decides the next step, and repeats until the task is complete.

Built-in tools (no configuration needed):

read_file — read any file from the working directory
write_file — create or overwrite files
list_directory — enumerate folder contents
file_exists — check before reading
run_command — execute shell commands

Three ways to add more tools:

Connect downstream workflow nodes as tools (available on AI Autonomous Agent and AI Agent Native Node — not on the standard AI Agent Node)
Load tools from the Tool Library (~/.nora/tools/)
Use the built-in file system tools as-is

Safety controls:

Max iterations: Default 10, configurable up to 100. The agent stops after this many tool-call cycles regardless of completion state.
Timeout: Default 30 minutes, configurable up to 120 minutes.
Budget limit in USD. Default $10. Set a dollar cap on API costs. The agent stops when the budget is reached.
Output routing: The agent’s final state routes to one of four paths — complete, partial, needs-input, or error. Downstream nodes handle each case differently.

SSE streaming: Watch the agent’s reasoning and tool calls in real time via the dashboard.

Upstream context merging: The agent automatically receives output from upstream nodes as additional context, so it builds on work already done in the workflow.

When to use it: Tasks with unpredictable step counts — “analyze these 50 CSV files and write a summary report,” “refactor all Python files in this directory to use type hints,” “read the error logs and fix the configuration.”

Level 3: AI Agent Native Node — Provider-Native Tool Calling

The AI Agent Native Node is an autonomous agent that uses the provider’s native tools / tool_use API parameter rather than parsing JSON from prose responses. The model is structurally required to return a tool call — it cannot return raw text or shell scripts in place of a structured tool call. This produces more reliable, consistent execution compared to JSON-in-prompt parsing.

Key differences from the AI Autonomous Agent:

Structured tool calls enforced at the API level — no fragile JSON extraction from model output
Output routes: complete, partial, and error (no needs-input edge in this version)
Default provider: Anthropic (claude-sonnet-4-6), though OpenAI and Google are supported
Same safety controls: configurable max iterations (default 10, up to 100), timeout (default 30 min, up to 120 min), and budget limit (default $10)
SSE streaming for real-time visibility into the agent’s reasoning and tool calls
Tool Library access with optional fuzzy discovery

When to use it: When you need reliable, structured tool execution and the standard autonomous agent’s JSON parsing is producing inconsistent results. Also useful when the strict contract between the LLM and its tools matters more than flexibility.

Level 4: Custom Script Agent — Write Your Own Agent Logic

The Custom Script Agent Node lets you write the agent loop yourself in Python or Node.js. Your script communicates with NORA through a JSON stdin/stdout protocol.

Your script receives instructions via stdin, calls its own LLM (with its own API keys), and sends tool_call requests back to NORA when it needs to execute tools. NORA runs the tools and returns results to the script’s stdin. The script repeats until done.

Key capabilities:

Any language model. Your script manages its own API keys and can call any provider — not limited to the three built into NORA.
Dynamic Tool Discovery. Your script can query NORA’s Tool Library with fuzzy search to find available tools at runtime.
Dynamic routes. The script decides its own output route labels, creating flexible downstream branching.
SSE streaming. Stream progress updates back to the NORA dashboard in real time.
Debug mode. Verbose logging of every JSON message between the script and NORA.

When to use it: When you need full control over the agent loop — custom retry logic, specialized prompting strategies, integration with APIs that aren’t in the Tool Library, or agent architectures that don’t fit the built-in autonomous pattern.

Cost Tracking Across All Agent Types

Every AI node in NORA tracks costs in real time. Per-call token counts, per-provider pricing rates, per-workflow totals, and per-tool breakdowns are all recorded in execution history.

The Autonomous Agent’s budget limit makes this actionable — set a $2.00 cap on an experimental workflow and it stops before overspending.

Practical Use Cases

Document processing pipeline. An AI Autonomous Agent reads a folder of mixed-format files, classifies each one, extracts key data, writes structured output to a results directory. If a file can’t be parsed, the agent routes to the error output for manual review.

Code generation with file I/O. Give an Autonomous Agent a spec file and tell it to generate the implementation. The agent reads the spec, writes code files, runs tests via run_command, reads test output, fixes failures, and repeats until tests pass — all within a configurable iteration and budget limit.

Data analysis with tool chains. A Custom Script Agent receives a dataset path, calls a Python analysis library through NORA’s tool system, reads the statistical output, decides which visualizations to generate, and produces a final report — using whatever LLM provider your script is configured for.

Getting Started

Download NORA at software.reibuys.com/nora. Install the .exe on Windows 10 or later. A paid license key is required. Drag an AI Agent Node onto the canvas, connect your API key for OpenAI, Anthropic, or Google, wire up downstream tools, and run it.

One-time purchase. No subscription. 30-day money-back guarantee.

Feature	AI Agent Node	AI Autonomous Agent	AI Agent Native Node	Custom Script Agent
LLM decides tool calls	Yes	Yes	Yes	Yes (your script)
Multi-turn conversation	Yes	Yes	Yes	Your implementation
Built-in file system tools	No	Yes (5 tools)	Yes (5 tools)	Via Tool Library
Tool Library access	Via connections	Yes	Yes + fuzzy discovery	Yes + fuzzy search
Budget limit	No	Yes (USD cap)	Yes (USD cap)	Your implementation
Max iterations	Max turns config	100 cap	100 cap	Your implementation
Timeout	N/A	120 min max	120 min max	Your implementation
Custom agent logic	No	No	No	Full control
SSE streaming	No	Yes	Yes	Yes
Provider-native tool calling	No	No	Yes	Your implementation
Output routing	Based on LLM choice	4 routes	3 routes (complete/partial/error)	Dynamic routes

Get NORA at software.reibuys.com/nora. One-time purchase. 30-day money-back guarantee.