Skip to content

Agent Frameworks Track

This track teaches agentic systems through a public-safe progression: deterministic local workflows first, optional hosted SDKs second, and reproducible evidence throughout.

  1. projects/autoresearch
  2. projects/agentic-course-assistant-showcase
  3. projects/modern-nlp-pipeline-showcase
  4. projects/model-release-rollout-showcase
flowchart LR
    A["Autoresearch<br/>agent-guided research loop"] --> B["Agentic Course Assistant<br/>tools, guardrails, traces, evals"]
    B --> C["Modern NLP Pipeline<br/>language pipeline mechanics"]
    C --> D["Model Release Rollout<br/>deployment evidence and rollback"]

Core Skills Covered

  • Separating deterministic workflow logic from model-backed agent behavior.
  • Routing questions to specialist intents.
  • Turning course resources into tool outputs that can be inspected.
  • Reading traces for route, tool, guardrail, and artifact evidence.
  • Comparing OpenAI Agents SDK and Google ADK concepts without requiring live credentials for the default path.
  • Designing eval rubrics before adding agent-as-judge scoring.
  • Deciding when memory, A2A, MCP, hosted tools, sessions, callbacks, plugins, and deployment are actually needed.

Primary Showcase

Start with projects/agentic-course-assistant-showcase.

cd projects/agentic-course-assistant-showcase
make sync
make smoke
make verify

Then read:

Evidence Artifacts To Inspect

  • projects/agentic-course-assistant-showcase/artifacts/course_assistant_response.md
  • projects/agentic-course-assistant-showcase/artifacts/agent_trace.json
  • projects/agentic-course-assistant-showcase/artifacts/resource_matches.csv
  • projects/agentic-course-assistant-showcase/artifacts/concepts/agentic_concepts.csv
  • projects/agentic-course-assistant-showcase/artifacts/concepts/openai_vs_adk_concepts.json
  • projects/agentic-course-assistant-showcase/artifacts/evals/agent_judge_rubric.json
  • projects/agentic-course-assistant-showcase/artifacts/evals/concept_coverage.json

Optional Live Framework Extension

The live path is intentionally opt-in:

  1. Run the offline smoke path and verify artifacts.
  2. Install only the optional SDK extra you need.
  3. Configure credentials outside source control.
  4. Run the optional reference module locally.
  5. Compare live behavior with the deterministic trace.

OpenAI Agents SDK:

cd projects/agentic-course-assistant-showcase
make sync-openai

Google ADK:

cd projects/agentic-course-assistant-showcase
make sync-adk
uv run adk run adk_course_assistant

Do not add live SDK execution to default CI. The public classroom contract is the offline harness and its artifact verifier.

flowchart TD
    Offline["Offline harness"] --> Evidence["Stable artifacts"]
    Evidence --> Verify["make verify"]
    Verify --> Optional{"Optional live extension?"}
    Optional -- "OpenAI" --> OAI["OpenAI Agents SDK<br/>tools + handoffs + tracing"]
    Optional -- "Google" --> ADK["Google ADK<br/>root_agent + tools + sessions"]
    OAI --> Compare["Compare against offline trace"]
    ADK --> Compare
    Compare --> CI["Default CI stays offline"]

Suggested Reflection Prompts

  • Which parts of the assistant should remain deterministic even after adding a hosted model?
  • What does the trace prove that the final answer alone does not prove?
  • When should a specialist be a handoff instead of a tool?
  • Which guardrail belongs in code rather than a prompt?
  • What would convince you that memory improved the experience without leaking private data?
  • What is the smallest eval dataset that would catch a routing regression?
  • Which live SDK feature is worth adding first, and which should stay out of the first build?