The response was lost. Was the message sent?
A practical guide to uncertain outcomes, idempotency keys and safe retries, with a worked transactional design and explicit limits on external side effects.
Insights
Technical articles on distributed systems and AI, with practical examples and sources you can check.
Memory is not authority.
How to separate remembered information from permission to act: provenance, scoped approvals, host-side enforcement and safe tool execution in AI systems.
Selected sources
Lectures, standards and engineering perspectives for further reading.
Threads and RPC
MIT · 6.5840. A concise foundation for reasoning about remote calls, concurrency and the uncertainty left by a missing response.
Making retries safe with idempotent APIs
AWS · Builders’ Library. A production-informed discussion of request identifiers, caller intent and the limits of deduplicating repeated operations.
HTTP Semantics
IETF · RFC 9110. The normative reference for HTTP methods, safety and idempotency. Useful for separating protocol guarantees from application policy.
The Protection of Information in Computer Systems
MIT · Saltzer & Schroeder. Foundational security principles, including least privilege and complete mediation, that remain useful when designing tool boundaries.
Zero Trust Architecture
NIST · SP 800-207. A framework for making access decisions around resources and explicit policy rather than assumed trust from location or history.
Effective context engineering for AI agents
Anthropic · Engineering. Practical context-management patterns, including compaction and external notes. Read alongside independent authorization controls.
Architecture and host responsibilities
Model Context Protocol. A versioned description of hosts, clients and servers. A useful vocabulary for locating control and consent around tools.