Capability Security
Running untrusted code without granting it the machine.
4 articles by Paul Scarrone.
- Agentic AI Engineering: Building Determinism Around a Non-Deterministic Engine A working definition of agentic AI engineering: treating tool-using language models as processes rather than conversations, and putting the determinism in the pipeline instead of the prompt.
- Orchestrating Agents on Amazon Bedrock: When You Cannot Push, Make Completion the Wake Multi-agent designs assume push delivery, which Bedrock-hosted sessions cannot receive. Making process completion the wake signal, and scoping agent credentials with session policies so blast radius is a property of the credential rather than the prompt.
- Five Techniques That Keep Paying Off When You Build for LLMs Six months of building tools for LLMs, distilled into five techniques: embedding and vector search, structural indexing, harness engineering, hooks, and why agents are better at writing code than using tools.
- Beamlet A minimum viable OTP over WebAssembly - supervised, capability-sandboxed processes on wazero, built as the substrate for a self-rewriting LLM harness