Skip to content

Examples

28 example workflows covering basic patterns to production-grade pipelines. All examples run with Ollama (free, local) or any cloud provider.

# Validate any example
agentloom validate examples/01_simple_qa.yaml

# Run with Ollama
agentloom run examples/01_simple_qa.yaml --provider ollama --model phi4

# Run with OpenAI
export OPENAI_API_KEY=sk-...
agentloom run examples/02_chain_of_thought.yaml

# Visualize the DAG
agentloom visualize examples/03_router_workflow.yaml
agentloom visualize examples/03_router_workflow.yaml --format mermaid

Basic

01 — Simple QA

Single LLM call. Sends a question, gets an answer.

Demonstrates: basic workflow structure, state, output mapping.

name: simple-qa
config:
  provider: ollama
  model: phi4
state:
  question: "What is Python in one sentence?"
steps:
  - id: answer
    type: llm_call
    prompt: "Answer this question concisely: {state.question}"
    output: answer

02 — Chain of Thought

Three sequential LLM calls: break down topic, research subtopics, synthesize summary.

Demonstrates: sequential dependencies, state passing between steps.

03 — Customer Support Router

Classifies user intent, routes to specialized handler (billing/technical/general).

Demonstrates: router step, conditional branching — only one branch executes.

04 — Tool Augmented

Fetches data from a URL using the http_request tool, then analyzes it with an LLM.

Demonstrates: tool integration, mixing tool and llm_call steps.


Intermediate

05 — Content Moderation Pipeline

Parallel content moderation for UGC platforms. Runs toxicity, PII, and policy checks simultaneously, aggregates results, and routes to approve/review/reject.

Demonstrates: parallel execution (3 checks at once), aggregation, router with 3 branches.

06 — Lead Qualification

B2B lead qualification pipeline. Enriches company data via API, analyzes buying intent, scores the lead, and routes to personalized outreach, nurture, or archive.

Demonstrates: parallel tool+LLM execution, multi-step scoring, conditional routing.

07 — Incident Triage

Automated incident triage for SRE teams. Fetches deployment context, correlates with alert data, classifies severity (P1/P2/P3), and generates response actions.

Demonstrates: 5-layer deep pipeline, tool integration, severity-based routing.

08 — Contract Risk Analysis

Legal contract risk analysis. Extracts clauses, evaluates risk, compares against industry standards, and generates an executive summary with sign/negotiate/walk-away recommendation.

Demonstrates: deep 4-step sequential chain, state accumulation, structured analysis.


Advanced

09 — Fraud Detection Pipeline

E-commerce order fraud detection. Fetches order + customer data in parallel, runs 4 concurrent fraud signal checks (velocity, amount, address, device), aggregates risk, and routes high-risk orders to a deep investigation subworkflow.

Demonstrates: 2 parallel tools, 4 parallel LLM checks, subworkflow, 11 steps / 5 layers.

10 — Multi-Market Content Localization

SaaS content localization across 3 markets. Fetches source content, runs 3 parallel localization subworkflows (ES/DE/JA) — each with 4 steps: translate, culturally adapt, legal compliance, and format — then cross-market review and deployment manifest.

Demonstrates: 3 parallel subworkflows (12 nested steps), 18 total steps across 4 layers.

11 — Insurance Claims Processing

End-to-end claims adjudication. Extracts structured data, runs 3 parallel validation tracks via subworkflows (coverage, fraud, medical necessity), aggregates results, routes to auto-approve, adjuster assignment, or denial.

Demonstrates: 4 subworkflows, 6 layers, 21 total steps. The most complex example.


Built-in Tools

12 — Log Analysis & Alerting

Server log analysis with shell_command tool. Collects error logs and system metrics in parallel, correlates them with an LLM, classifies severity, and routes to PagerDuty alert, ops ticket, or silent log.

Demonstrates: shell_command tool, parallel data collection, severity routing.

13 — Report Generator

Data pipeline health report using file_read and file_write tools. Reads pipeline results and SLA config, analyzes compliance, generates executive summary, and writes the final report to disk.

Demonstrates: file_write + file_read tools, file-based data pipeline.

14 — Custom Tools: @tool Decorator

Sentiment monitoring pipeline with custom tools defined using the @tool decorator: query_database, send_slack_message, create_ticket.

Demonstrates: @tool decorator, custom tool registration.

uv run python examples/14_custom_tools_decorator.py

Two files

This example uses a paired YAML workflow + Python runner: 14_custom_tools_decorator.yaml + 14_custom_tools_decorator.py

15 — Custom Tools: BaseTool Subclass

Customer data enrichment with custom tools defined as BaseTool subclasses: GeocodingTool, CRMLookupTool, RiskScoreTool.

Demonstrates: BaseTool subclass pattern, churn risk scoring.

uv run python examples/15_custom_tools_subclass.py

Resilience & Sandbox

16 — Circuit Breaker Demo

Intentionally trips the circuit breaker with a non-existent model, then shows automatic fallback.

Demonstrates: circuit breaker state transitions (CLOSED -> OPEN), provider fallback.

17 — Sandbox: Allowed Operations

Runs commands and file I/O within sandbox limits. echo is allowed, files stay inside /tmp/agentloom, network is enabled.

Demonstrates: config.sandbox, command allowlist, path restriction.

18 — Sandbox: Blocked Operations

Attempts operations that violate sandbox policy: disallowed command (curl), path outside allowed directory, network disabled, and pipe injection (echo | cat).

Demonstrates: sandbox enforcement, command injection prevention, path and network blocking.


Multi-modal

19 — URL Image

Fetches an image from a public URL and describes it. The engine downloads the image locally and sends base64 to the provider (fetch: local).

Demonstrates: attachments, URL fetch, multi-step vision pipeline.

20 — Base64 Inline

Analyzes an image embedded directly in workflow state as base64. No network access needed.

Demonstrates: inline base64 attachment, offline-capable vision.

21 — URL Passthrough

Sends the image URL directly to the provider API (fetch: provider). Only works with OpenAI and Anthropic.

Demonstrates: fetch: provider mode, provider-side image fetching.

22 — Sandboxed URL

Image analysis with sandbox restrictions. Only domains in allowed_domains are permitted.

Demonstrates: sandbox allowed_domains for attachments.

23 — PDF Document

Extracts key points from a PDF and generates an executive summary. Requires Anthropic or Google.

Demonstrates: type: pdf attachment, document analysis.

24 — Audio Transcription

Transcribes an audio clip and analyzes for topic, sentiment, and action items. Requires OpenAI or Google.

Demonstrates: type: audio attachment, transcription + analysis pipeline.


Streaming

25 — Streaming QA

Streams LLM output token-by-token in real-time.

Demonstrates: stream config, --stream CLI flag, time-to-first-token tracking.

agentloom run examples/25_streaming_qa.yaml --stream

26 — Streaming + Multi-modal

Combines streaming with image input. The image is fetched locally, and the LLM description is streamed back in real-time.

Demonstrates: streaming + attachments composability.


State Features

27 — Array Index

Array index support in state paths (state.items[0], items[0].name, results[-1]).

Demonstrates: bracket-based array indexing in state and templates.


Checkpointing

28 — Checkpoint & Resume

Two-step workflow with checkpointing enabled. Persists execution state so failed or interrupted runs can be resumed without re-executing completed steps.

Demonstrates: --checkpoint flag, agentloom runs, agentloom resume.

# Run with checkpointing
agentloom run examples/28_checkpoint_resume.yaml --checkpoint --lite

# List checkpointed runs
agentloom runs

# Resume a failed or interrupted run
agentloom resume <run_id> --lite

Testing & Replay

31 — Record and Replay

Two-step workflow used to capture real LLM responses and replay them offline. Run once with --record against a real provider, then replay any number of times without network or API keys.

Demonstrates: --record, --mock-responses, agentloom replay.

# Capture real Anthropic responses (per-call flush; partial recordings survive crashes)
agentloom run examples/31_record_and_replay.yaml --record recordings/byzantine.json

# Replay offline — pick either form
agentloom replay examples/31_record_and_replay.yaml --recording recordings/byzantine.json
agentloom run examples/31_record_and_replay.yaml --mock-responses recordings/byzantine.json

32 — YAML-configured MockProvider

Same workflow as 31, but with provider: mock and responses_file declared in YAML. Runs via plain agentloom run with no CLI flags — useful for committed fixtures and CI.

Demonstrates: provider: mock, responses_file, latency_model: replay.

# Depends on the recording captured from example 31
agentloom run examples/32_yaml_mock.yaml --lite