4 min

Workflows

Multi-Step Automation Pipelines

From Single Tools to Pipelines

A single tool does a single task. A workflow chains multiple tools into an automated pipeline — scan your downloads, organize by type, summarize new documents, and send a daily report. The real power of agents comes from this composition.

But chaining actions multiplies risk. If step 1 moves a file and step 3 fails, you might lose track of where things are. Workflows need error handling, rollback logic, and safety controls that single tools don't.

Workflow Architecture

A workflow is a directed graph of steps, where each step is a tool call with inputs, outputs, and error handling:

Step 1 (scan) → Step 2 (filter) → Step 3 (process) → Step 4 (report)
                     ↓ (no files)          ↓ (error)
                   [skip]              [rollback + alert]

Each step has:

Component	Purpose
Input	Data from the previous step or initial parameters
Action	The tool call that does the work
Output	Data passed to the next step
Error handler	What to do if this step fails
Rollback	How to undo this step if a later step fails

Error Handling Strategies

Different errors need different responses:

Error Type	Example	Strategy
Transient	API timeout, rate limit	Retry with exponential backoff (max 3 attempts)
Permanent	File not found, permission denied	Log error, skip item, continue workflow
Critical	Safety rule violation, budget exceeded	Halt workflow immediately, alert user
Partial	8 of 10 files processed, 2 failed	Complete what's possible, report failures

Never silently swallow errors. Every failure should be logged and reported, even if the workflow continues.

Rollback Logic

Rollback answers the question: "If step 3 fails, how do I undo steps 1 and 2?"

Not every action is reversible. Here's how to handle common cases:

File moves — record original and destination paths, move back on rollback

File creation — record created file paths, delete on rollback

API calls — generally not reversible — use compensating actions instead

Data modifications — create backups before modifying, restore on rollback

The key principle: record enough state before each step to undo it later.

Workflow Safety Controls

Workflows need additional safety beyond what individual tools provide:

Timeouts — set a maximum execution time for the entire workflow. A workflow that runs forever is a workflow that's stuck.

Max retries — limit how many times a failed step can retry. Without limits, a retry loop becomes an infinite cost generator.

Budget tracking — track cumulative API costs across all steps. A workflow with 10 steps that each costs $0.05 adds up to $0.50 per run — which adds up to $15/month if it runs hourly.

Human-in-the-loop — some actions should require explicit user approval:

Deleting more than N files

Spending more than $X in a single workflow

Sending external API calls or emails

Modifying files outside the working directory

Example: File Management Workflow

Here's a practical 3-step workflow:

Step	Tool	Input	Output	On Error
1. Scan	file-organizer	downloads.json	categorized file list	Halt — can't proceed without files
2. Organize	file-organizer	categorized list	move results	Rollback moves, report
3. Report	report-generator	move results	summary document	Log warning, workflow still succeeds

Notice that step 3 failing doesn't require rolling back steps 1-2 — the files are already organized correctly. Error handling should be proportional to the impact.

Testing Workflows

Test each layer independently:

Unit test each tool in isolation

Integration test the full workflow with sample data

Error injection — force individual steps to fail and verify rollback

Budget test — verify cost tracking across all steps

Timeout test — verify the workflow halts when the time limit is reached

Key Takeaways

Workflows chain tools into pipelines — powerful but riskier than single tools.

Every step needs an error handler and rollback logic. Partial completion is often worse than none.

Use different strategies for transient, permanent, critical, and partial errors.

Add timeouts, max retries, budget tracking, and human-in-the-loop for high-risk actions.

Test workflows with error injection — force failures and verify the system recovers correctly.

This is chapter 5 of Open Source AI Agents (OpenClaw).

Get the full hands-on course — free during early access. Build the complete system. Your projects become your portfolio.

View course details

Ch. 4: Custom Tools

Ch. 6: Monitor & Maintain