Agent Infrastructure: Memory, Access, and Harness Patterns

Building an Advanced Agentic Harness lays out how to compose small, testable primitives into a production-ready agent harness that plans, executes, verifies, and budgets multi-step LLM workflows. If you’re building outcome systems, the post is a hands-on blueprint for reliable orchestration, verification, and cost control across agentic pipelines (Principles 09 & 13).

Zero-Mem: Zero-Token Memory Operations for LLM Agents presents a deterministic retrieval approach using entity graphs and temporal hierarchies to remove LLM token and call costs for agent memory. That design dramatically cuts run and call costs while preserving structured long-term context—useful when you need agents to hold state without exploding token budgets (Principles 02 & 06).

The Agent Access Model proposes enforcing least-privilege, short-lived capabilities for machine agents operating at machine speed. Treat this as a security pattern for agents: design short-lived, capability-limited credentials into your orchestration layer to shrink blast radius and enable automated revocation (Principles 10 & 15).

WriteGuard: fine-grained controls for MCP Servers launches a control plane to centrally control, attribute, and audit agent write actions on MCP servers to prevent runaway automated changes. Add write-level gates and attribution to your deployment paths so agent-driven changes are auditable and can be stopped before they become incidents (Principles 10 & 15).

llm-anthropic 0.26 adds Claude 5 support, server-side tools, and streaming typed events via llm 0.32. Server-side tools and typed streaming events make agents more composable and observable—integrate them into your runtime to improve legibility and create event graphs you can route, monitor, and verify (Principles 06 & 11).