The complete agent development curriculum — same concepts, same depth, zero API costs. Run everything locally with Ollama or for free on Groq.
🔍 Search the Course · 💬 Ask the Course (works offline)
On a phone or tablet? Try the mobile-optimized versions — condensed cards and pseudocode for learning on the go.
View Mobile Versions ›Before starting any hands-on module, complete the Dev Environment Setup module. It walks you through installing Python, VS Code or PyCharm, creating a virtual environment, installing the openai package, and getting Ollama + Mistral-7B running locally. Ends with a working verification script. Takes 30–40 minutes and only needs to be done once.
Everything else — agent architectures, RAG pipelines, guardrails, multi-agent patterns — is identical. The concepts are model-agnostic.
STEP 0 — START HERE
STEP 0B — HELLO WORLD
TRACK 1 — FOUNDATIONS
How language models work from first principles — attention, context windows, tokens. First API call with Mistral-7B via Ollama.
Zero-shot, few-shot, chain-of-thought. System prompts, role prompting, structured output with Mistral.
Getting reliable JSON from open source models. Response validation, error recovery, Pydantic schemas.
How tokenizers work, context limits on open source models, prompt budgeting. Count tokens with tiktoken and Mistral's tokenizer before making API calls.
What the model actually sees: six layers of context. Four levers (crop, compress, summarize, select). Position effects — lost in the middle. Context rot. Build a ContextBudget class for local models with tight windows.
TRACK 2 — TOOL USE
Tool use with Mistral — define tools, parse tool calls, handle results. The foundation of agentic behavior.
Parallel and sequential tool chains. Let Mistral pick the right tool, chain outputs, handle failures.
The Model Context Protocol — USB-C for agents. Build an MCP server with FastMCP, bridge it into a local Mistral agent (MCP tools → OpenAI tool_calls), add AI-native web search, and produce rich visual output: a tool that renders a chart and returns it as an image artifact.
Give your Mistral agent the ability to run code. subprocess sandbox with resource limits, Docker-isolated execution, self-debugging loop. The agent writes code, runs it, reads the output, and fixes errors — all locally.
TRACK 3 — MEMORY & CONTEXT
Managing history across turns with a stateless API. Sliding window, summarization, token budget allocation.
Build a local RAG pipeline. Embed with sentence-transformers, store in ChromaDB, retrieve and generate with Mistral — no external APIs.
Chunking strategies, hybrid search (BM25 + dense), re-ranking with a local cross-encoder, multi-query expansion. Production-grade retrieval pipelines without any cloud dependency.
In-process buffer, semantic vector store (ChromaDB), episodic memory (summarization). Build an agent that remembers across sessions — all locally, no external service.
TRACK 4 — AGENT ARCHITECTURES
Build the Reason→Act→Observe loop with Mistral. Manual implementation — no frameworks. See exactly how the agent loop works.
Intent classification, DAG execution, dynamic tool discovery. Build a planning agent that breaks complex tasks into steps.
Supervisor/worker, peer-to-peer, pipeline architectures. Build a content pipeline with multiple Mistral agents passing messages.
TRACK 5 — GUARDRAILS & SAFETY
PII detection, prompt injection defense, schema validation. These guardrails protect any LLM agent — Mistral or Claude.
Hallucination detection, cost controls, approval gates. Circuit breaker pattern for when your local model goes off-rails.
TRACK 6 — EVAL & OBSERVABILITY
Build an eval harness for your Mistral agents. Automated test suites, LLM-as-judge with a local evaluator model, DeepEval + Ragas for RAG pipelines. Zero cloud dependency — Ollama as judge.
Structured traces with LangSmith and OpenTelemetry. Works with any model — add trace IDs, log tool calls, visualize multi-agent execution chains.
Production dashboards for latency, error rates, token spend per agent. Detect model drift, trigger re-evals, close the feedback loop from real traffic.
TRACK 7 — PRODUCTION
Wrap your Mistral agent as a FastAPI service. Request queuing, async execution, streaming responses, health checks. Deploy with Docker — no vendor lock-in.
Model selection (7B vs 13B vs 70B trade-offs), quantization (Q4 vs Q8), batch inference, prompt caching patterns, token budgeting. Minimize cost without sacrificing quality.
Deploy Ollama to GPU cloud VMs on GCP, AWS, and Azure. Cloud-hosted open source alternatives: Amazon Bedrock (Llama), Vertex AI Model Garden, Together AI, Groq. Build a provider-agnostic wrapper — swap local for cloud with one env var.
Run agents with no UI and no human in the loop — driven by cron, CI, or a webhook and consumed by another program. The headless contract (JSON on stdout, logs on stderr, meaningful exit codes), guardrails that replace the human circuit-breaker, and Unix-pipeline composition.
Engineering scaffolding that makes an agent a safe contributor: reusable workflow files, a spec→generate→verify framework, structural linting (ruff, bandit, ast-grep) that beats regex, and pre-commit + CI gates that turn a failing check into a correction loop the agent fixes itself.
TRACK 8 — WHAT'S NEXT
claude-agent-sdk is Anthropic-only. Use M11–M13 to build the same patterns manually, or use CrewAI (Capstone C4).
ollama pull deepseek-r1).
Once you've completed all 26 modules, you'll understand the core agent development loop end-to-end. If you want to progress to Claude-specific features (prompt caching, extended thinking, hooks, the Agent SDK), the main Claude Agents course builds directly on this foundation — your agent code is compatible, just swap the client back.
Continue to the Claude Agents course →Three curated paths through the 26-module curriculum. All paths use Ollama/Mistral locally — zero API cost.
Want to continue to Claude-specific features after this track? The main Claude Agents course builds directly on this foundation — your agent code is compatible, just swap the client.
3 progressive capstones across 2 industry domains. Complete the prerequisite track modules before starting each one. All run 100% locally on Ollama — no API cost.
After M13 + M16 + M17 · 4–5 hours · CrewAI · Coming soon
After M16 + M18 · 4–5 hours · DeepEval · Coming soon
The main Claude Agents course has 7 capstone tiers × 3 domains = 19 capstone labs. This open source track has 3 capstones covering the same progression (autonomous agent → multi-agent → eval-driven) but adapted for Ollama/Mistral and open-source frameworks (CrewAI, DeepEval). Capstone 7 of the main course (spec-driven via Claude Code) has no open-source equivalent.
Module dependencies — complete prerequisites before moving on.