Archive
Written for CrabTalk's earlier agent runtime, and kept as written.
- L0 to L5: where AI agents land on the autonomy spectrum
- MiroFish: swarm prediction through social simulation
- Async compaction: the race conditions nobody talks about
- Built-in agents: what ships in the box
- Built-in tools: what your agent can reach
- Hermes memory: five layers, one learning loop
- Mem0: what three memory scopes actually cost
- Hermes Agent: what Nous Research built
- When agents share a channel
- Built-in web search: no API keys, no setup
- Context compaction in agent frameworks
- Tool permissions and the bash bypass problem
- How agents call agents
- SOUL.md: brilliant idea, brittle implementation
- How AI frameworks control model thinking
- What should an agent capability bench test?
- How developers configure their AI agents
- Sandboxing AI agents: beyond Docker and WASM
- Why multi-agent workflows fail in production
- How AI agents remember: a survey of persistent memory
- Plans vs tasks: how AI agents think before they act