AgentOS Reader Resources

Complete figure PDF · 24.3 MiB SHA-256

Illustrations are hosted on this site. Complete Cloudflare verification before downloading. PDFs are moderately compressed and link back to the author's website from the cover and every page footer. The collection contains illustrations, figure numbers and captions, without the book's body text.

Download companion source ZIP Download SHA-256 file

GitHub: AgentOS-Lab companion source · Report an issue · Repository code is available under the MIT license.

Size: 373285 bytes (about 365 KiB). Includes a usage guide and one agentos_lab project.

SHA-256: 495c230f5cddedb29dec7e09f0a803dd997191ee8595a68af936c0887c0287c5

The archive uses the earlier title 《从零搭建企业级 AgentOS》. This is the original delivery archive; this page uses the published title. Follow the directory structure and usage guide inside the download.

1. Run the first Agent

Uses only the Python standard library; prints a simulated purchase recommendation without writing to an ERP.

python agentos_lab/agentos_lab/first_agent.py

2. Install dependencies and run the purchase case

Install uv first. The initial dependency installation requires internet access.

uv sync --project agentos_lab --locked
uv run --project agentos_lab python -m agentos_lab.main run purchase --case normal --json

Default cases need no model API key or enterprise account. Companies, ERP, OA and business data are simulated. Real models, remote MCP, OPA, Temporal and telemetry backends require extra configuration; framework commands show local interface mappings.

3. Explore long-running task recovery

python agentos_lab/run_purchase_task.py --task purchase-demo
python agentos_lab/send_approval.py --task purchase-demo
python agentos_lab/workers/purchase_worker.py
python agentos_lab/query_task.py --task purchase-demo

Run these commands in order from the same directory; they share .agentos-demo/tasks.sqlite3. The worker processes the approved task and the query should return completed. Repeated processing does not create duplicate simulated purchase records. Use a new task ID for a new exercise.

1. 1. The first Agent

   • agentos_lab/first_agent.py

   python agentos_lab/agentos_lab/first_agent.py

   Standard library only; prints a simulated purchase recommendation.

2. 2. Object contracts

   • agentos_lab/domain.py
   • agentos_lab/schema_export.py
   • object_contract/

   uv run --project agentos_lab python -m agentos_lab.main framework schemas --json

   Exports core object contracts; object_contract/ contains a separate handwritten validation example.

3. 3. Task runtime

   • agentos_lab/runtime.py
   • agentos_lab/graph_compat.py
   • agentos_lab/langgraph_api.py

   uv run --project agentos_lab python -m agentos_lab.main run purchase --case write-timeout --json

   Exercises a purchase timeout locally; LangGraph requires the optional runtime dependencies.

4. 4. Model gateway

   • model_gateway/
   • agentos_lab/model_router.py
   • agentos_lab/litellm_gateway.py

   uv run --project agentos_lab python agentos_lab/model_gateway/route_demo.py --case purchase_budget_limit

   Local budget fallback decision; does not call a model provider.

5. 5. Tool protocols

   • agentos_lab/fastmcp_adapter.py
   • agentos_lab/protocols.py
   • agentos_lab/connectors.py

   uv run --project agentos_lab python -m agentos_lab.main framework fastmcp --json

   Shows tool contracts and framework mappings; does not establish a remote MCP connection.

6. 6. Data and evidence

   • agentos_lab/knowledge.py
   • agentos_lab/data_engine.py
   • agentos_lab/llamaindex_context.py

   uv run --project agentos_lab python -m agentos_lab.main run document --json

   Runs the local document case; standalone context_evidence/ examples include preset fields.

7. 7. Policies and permissions

   • agentos_lab/policy.py
   • agentos_lab/control_plane.py
   • agentos_lab/opa_policy.py

   uv run --project agentos_lab python -m agentos_lab.main run purchase --case permission-denied --json

   Exercises permission denial; actual OPA execution requires a separately configured service.

8. 8. Traces and testing

   • agentos_lab/trace_export.py
   • agentos_lab/otel_observability.py
   • agentos_lab/evaluations.py
   • tests/

   uv run --project agentos_lab python -m agentos_lab.main trace purchase --case write-timeout --json

   Inspects local execution traces without a telemetry backend.

9. 9. Long-running task recovery

   • runtime_queue/local_queue.py
   • run_purchase_task.py
   • send_approval.py
   • workers/purchase_worker.py
   • query_task.py

   python agentos_lab/run_purchase_task.py --task purchase-demo

   Submit first, then approve, process and query using the recovery walkthrough below; uses local SQLite without Temporal. Full recovery walkthrough

10. 10. System composition

    • agentos_lab/system.py
    • agentos_lab/registry.py
    • composition/

    uv run --project agentos_lab python -m agentos_lab.main run all --json

    Runs core scenarios; preset composition/ and composition_walkthrough/ outputs are not integration verification.

11. 11. Purchase and equipment cases

    • agentos_lab/runtime.py
    • agentos_lab/operations_runtime.py
    • agentos_lab/casebook.py

    uv run --project agentos_lab python -m agentos_lab.main run service --case stale-knowledge --json

    Exercises stale equipment documentation and blocks writeback; use run purchase --case erp-conflict --json for purchase conflicts.

12. 12. Platform operations

    • platform_registry/
    • agentos_lab/backstage_catalog.py
    • agentos_lab/benchmark.py

    python agentos_lab/platform_registry/validate_catalog.py

    Checks required catalog markers rather than full JSON Schema validation; Backstage is an object mapping, not a running platform service.

2026-10-25 · Illustrations and public companion code

Added a complete PDF containing 89 illustrations from the final text, with figure numbers, captions and checksums. Companion code is now available in AgentOS-Lab for browsing and issue reports.

2026-09-10 · Integrated companion source

Brings the first Agent into a single project with a 12-chapter code guide, locked dependencies and local task recovery. The original delivery record reports 72 passing checks in each of the default and LangGraph environments on macOS arm64 with Python 3.10.16. This is a historical record, not verification of every operating system or real external service.

Errata and feedback

No confirmed errata entries have been collected here yet. Send the chapter, page, source version, command, error output and reproduction steps to xujiayu0606@gmail.com.

• MiniCode: building an AI coding assistant from scratch — Tool calling, context and permission boundaries.
• Architecture details of modern large language models — Model foundations behind the gateway.