The shift from "chat with an LLM" to "give an LLM tools and let it act."
This collection tracks the frameworks, patterns, and experiments around
autonomous AI agents.
vs. agents. When to use each. The Anthropic team's hard-won lessons.
to multi-agent systems. Each agent is a process. Supervision trees handle
failures. This is underexplored.
your machine. The privacy angle on agents.
listen and speak, not just read and write.
as an execution environment.
workflows. The meta-level: agents that build knowledge bases.
Agent = LLM + tools + loop. The LLM decides what to do, the tools
execute, the loop continues until done. The hard parts: knowing when
to stop, handling errors gracefully, not burning tokens on dead ends.
The Elixir/BEAM connection is interesting. The actor model was built
for exactly this problem: independent processes communicating via
messages, supervised by a hierarchy that handles crashes. Every AI
agent framework is reinventing Erlang/OTP.
as knowledge compilers
force multipliers for solo devs