Hermes Agent alternatives: personal agent or agent runtime?
Hermes Agent and OpenClaw are personal, self-hosted agents. When to use them, and when you need a runtime with A2A, tenancy and default-deny policy.
Blog
Security, memory, MCP and A2A: engineering notes on running AI agents in production.
FrançaisHow AI agent memory actually works: history, compaction, recall
Agent memory is three systems, not one. How context eviction, compaction thresholds and tenant-scoped recall work in a real runtime.
MCP vs A2A: the difference, with a working architecture
MCP connects an agent to tools. A2A connects callers to agents. How the two protocols fit together, and what to secure in each.
Securing an AI agent: enforce limits outside the model
Prompts are not a security boundary. Default-deny policy, command gating, SSRF refusal and sealed keys: how to constrain an agent at runtime.
Shipping your first production agent: a five-step checklist
Bound the loop, scope the tools, cap the spend, require sources, measure. A practical checklist for your first production AI agent.