Reference — examples by number

This page lists every shipped example by number with a one-line summary. Use it when you know the file number and want to jump straight to the source. For discovery by topic, start with Tutorials or Learn the SDK.

There are 47 shipped examples. The first 32 demonstrate individual SDK capabilities. The last 15 are full walkthroughs with sibling READMEs.

Quickstart

#	Example
01	hello_agent — `pip install && python` → working agent

SDK basics

#	Example	Capability
02	tool_agent	Tool calling
03	multi_model	Anthropic ↔ OpenAI swap
04	memory_agent	Memory basics
05	workflow	Multi-step workflows
06	guardrails	Safety filters
07	mcp_tools	MCP integration
08	directives	Directive library
13	model_routing	Routing rules
18	local_llm_routing	Ollama / LM Studio
27	app_factory	App-factory pattern

Integrations

#	Example	Tool
09	proxy_claude_code	Claude Code CLI
10	proxy_cursor	Cursor IDE
11	proxy_codex	OpenAI Codex
14	register_agent	Registry — publish
15	discover_agents	Registry — consume
23	harness_proxy	Harness proxy primitives
24	harness_agent	Harness agent integration

Production patterns

#	Example	Problem solved
12	budget_enforcement	Per-user/team/project budget caps
16	agent_governance	Approval flows + audit
19	domain_model	Domain-modelled agent state
20	fleet_workers	Distributed worker registration
21	full_stack	End-to-end SDK + workflow + fleet
22	memory_holds_across_llms	Memory portability across LLMs
25	training_data_pipeline	Curator-style data capture
26	fleet_scoped_dispatch	Capability-based dispatch
28	autopilot_quickstart	Autopilot without an LLM key
29	memory_strategies	Strategy-based memory extraction
31	grounded_multi_model	Multi-LLM grounded retrieval
32	global_shared_memory	Cross-agent shared knowledge
35	autopilot_hosted_service	Autopilot via hosted blueprint service

End-to-end walkthroughs

Each example below ships with a sibling README and maps to a tutorials page.

#	Example	Real-world problem	Tutorial
30	oncall_agent	Incident response	(cross-cutting)
33	fleet_sealed_integration	Multi-tenant production with isolated tenants	Production multitenancy
34	observatory_cost_tracking	Track AI spend by team and project	Observability and cost
36	autopilot_training_loop	Capture → fine-tune trigger	Train your own model
37	semantic_checkpoint_recall	Maintain long context across weak models	Memory and retrieval
38	unsloth_finetune	Real LoRA fine-tune, dual-backend	Train your own model
38a	mlx_lm_server_deploy	Apple Silicon deploy via `mlx_lm.server`	Inference deployment
39	sandbox_scoped_credentials	Sandbox + scoped credentials	Production multitenancy
40	fleet_under_load	20+ workers, live state HUD	Observability and cost
41	graph_memory_incident_dependency	Graph beats vector for multi-hop	Memory and retrieval
42	support_triage_agent	Customer-support triage	Moderation and classification
43	observatory_live	Real telemetry → live dashboards	Observability and cost
44	colab_free_cuda	Free Tesla T4 via Drive-sync	Inference deployment
45	vastai_marketplace_bid	Bid-cheapest with reliability scoring	Inference deployment
46	custom_inference_as_tool	Bring-your-own endpoint	Inference deployment
47	runpod_finetune_orchestration	RunPod reliable rental	Inference deployment
48	modal_serverless_inference	Per-second serverless	Inference deployment
49	community_moderation	Three classifiers + LLM judge	Moderation and classification

Notes

Example 17 — unsloth_finetune (legacy stub): superseded by Example 38 (real fine-tune). Marked deprecated in its docstring; left in place for backward compatibility.
No example numbered 02-08 in the _factory, _perf_*, _soaks, or notebooks subdirectories — those are dev tools, not user-facing examples.
Numbering gaps and sibling-suffixes (e.g. 38a) follow the example-build-conventions — when two PRs collide on a number, the later one gets a lowercase-letter sibling suffix to preserve numeric reservations.