Curriculum
What you learn, in order.
Sixteen tracks in five stages, from a first API call to agents that survive production. The first 4 are open now; the rest are coming soon. Jump to a stage, or scroll to walk the trail.
Start the curriculum next: 01 · the whole path
Foundations · 01 / 16
Model APIs
The model as a function you call, and everything that goes wrong when you do.
Start here 14 problems · 1 project · the list
What's inside
429s bounce off the wall and the agent backs off: 0.5s, then 1s, then 2s. A 400 shatters, and it is not retried.
messages and roles · sampling · stop reasons · tokens and pricing · streaming · retries and backoff · structured output
- a client that retries 429s but not 400s
- a parser for JSON a model wrapped in prose
Foundations · 02 / 16
Context engineering
What the model sees is all it knows. Decide what goes in, in what order, within a budget.
Start here 9 problems · the list
What's inside
Every turn pours more history into the window. Left alone it spills over the rim; compacted, the oldest turns press into a summary.
system prompts · few-shot examples · context budgets · ordering and recency · compaction · prompt caching · extended thinking
- a context builder that fits a token budget
- a history compactor that keeps what later turns need
Foundations · 03 / 16
Retrieval
Find the right text before you ask. Most bad answers are bad retrieval.
Start here 12 problems · the list
What's inside
The query is a beam. Cosine similarity is the angle between it and each document: the ones it lights up most are the top five the model reads.
embeddings · chunking · BM25 · hybrid fusion · reranking · HNSW · recall@k and nDCG
- hybrid search with reciprocal rank fusion
- the Production RAG build, scored on SciFact
Foundations · 04 / 16
RAG and grounding
Answers that cite what they used, and say so when nothing supports them.
Start here 10 problems · the list
What's inside
Each claim sends a thread to the passage words it shares. The claim whose thread finds nothing is flagged instead of shipped.
retrieve then generate · citations · groundedness checks · no-answer handling · multi-hop · retrieval as a tool
- a pipeline that returns citations
- a checker that flags unsupported claims
Agents · 05 / 16 · soon
Tool use
Let the model act. The schema and the description are the interface it programs against.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
The schema is a lock. Arguments that fit open the tool; the rest jam, and the exact error goes back to the model.
JSON Schema tools · descriptions that steer · argument validation · parallel calls · errors as results · side effects · tool search
- a dispatcher that validates arguments
- a parallel executor with per-tool timeouts
Agents · 06 / 16 · soon
The agent loop
Call, act, observe, repeat, and know when to stop.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
The agent laps the loop: call the model, run the tool, read the result, decide. When the model says end_turn, it takes the exit to the answer.
stop_reason control · turn and budget limits · message state · cancellation · approval gates · checkpoint and resume
- the loop itself
- a loop that stops inside a dollar budget
- a loop that resumes after a crash
Agents · 07 / 16 · soon
Workflows and planning
Most tasks want a workflow, not an agent. Know which, and build either.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
The orchestrator splits the work into three streams. They run at once, and the merge gate opens only when the slowest arrives.
prompt chaining · routing · parallel sections and voting · orchestrator and workers · evaluator and optimizer · plan then execute · reflection
- a router with a fallback
- an orchestrator that fans out and merges
Agents · 08 / 16 · soon
Memory and state
What survives the end of the context window, and who decides.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Every message falls onto a sieve. Chatter falls through; lasting facts stay. When two share a key, the newer overprints the older, and recall finds it.
scratchpads · conversation memory · long-term stores · write policies · forgetting and conflicts · per-user isolation
- a memory store with a write policy
- recall that prefers the newer of two facts
Agents · 09 / 16 · soon
Multi-agent systems
Split work to isolate context, and pay for every extra agent you add.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Each subagent fills its own bubble of context, so nothing leaks between them. Only small summaries float up to the supervisor, and every bubble is billed.
subagents · handoffs · supervisors · shared state · context isolation · fan-out cost · failure propagation
- a supervisor that delegates to subagents
- a handoff that carries only what the next agent needs
Connections · 10 / 16 · soon
MCP
One protocol to connect any tool or data source to any agent.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
One cable, one protocol. The client and server agree on capabilities, list the tools, then call one. Any server plugs in the same way.
hosts, clients, servers · tools, resources, prompts · stdio and HTTP transports · capability negotiation · auth
- an MCP server for a local dataset
- a client that discovers and calls its tools
Connections · 11 / 16 · soon
Coding and computer use
Agents that read a repo, edit it, run the tests, and click through a UI.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
The agent types a patch, runs the tests, reads what failed, and patches again, until every bulb is lit or the budget runs out.
repo navigation · applying patches · test-and-fix loops · sandboxed execution · browser actions · verifying outcomes
- a patch applier that survives drift
- a loop that runs tests until green or out of budget
Engineering · 12 / 16 · soon
Evals
If you cannot measure it, you are guessing. Grade outcomes and trajectories.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Each row is one episode of five tries at 80% success. pass@5 needs one head; pass^5 needs all five. The live estimates converge on the formulas.
outcome and trajectory evals · test sets · deterministic replay · LLM as judge · pass@k and pass^k · regression suites · benchmarks
- a trajectory grader
- a judge calibrated against human labels
Engineering · 13 / 16 · soon
Tracing and debugging
Read the run. Find the first turn that went wrong; the rest is downstream.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
A lens runs along both recordings frame by frame. At the first frame that differs it stops: that is the bug. Everything after is static.
traces and spans · GenAI semantic conventions · first divergence · replay · loop detection · tool error rates
- a trace differ
- a detector for an agent stuck repeating itself
Engineering · 14 / 16 · soon
Cost and latency
Correct is the starting line. Then get under the call, token and dollar budget.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Four tool calls, the same tokens either way. One after another, every hourglass waits for the last; in parallel, you wait only for the slowest.
token budgets · prompt caching · model routing · batching · parallel tool calls · context pruning · early exit
- the same agent at half the tokens
- a cascade that tries the small model first
Engineering · 15 / 16 · soon
Security
Every input is untrusted, including the ones your tools fetch.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Private data, an injected instruction, and a pipe to the outside. With all three open, data leaks. Close any one valve and nothing leaves.
prompt injection · the lethal trifecta · least-privilege tools · sandboxing · secrets · approval for irreversible actions · exfiltration
- a tool gate that blocks exfiltration paths
- a fix for an agent that obeys a poisoned web page
Production · 16 / 16 · soon
Shipping agents
Agents that survive real traffic, real failures and the next model release.
Coming soon. The problems for this track are being written. The first 4 tracks are open now.
What's inside
Both climbers fall at ledge 4. Without checkpoints one starts again from the ground and repeats three ledges; the other respawns at its last flag.
durable execution · queues and background runs · streaming UIs · prompt and model versioning · monitoring · fallbacks · feedback loops
- a run that resumes where it failed
- a rollout that compares two prompt versions