Filtered to AI Agents & Systems · show all articles
Five structural problems in agent design remain open by the field's own account — long-horizon credit assignment, durable state, multi-agent coordination cost, action verification, and silent sub-task failure.
A long agent loop resends its own history on every step, pays a fixed tool tax each time, and risks a retry that resends it twice. That arithmetic, not model quality, sets the real ceiling on how far a trajectory can run.
The failure is rarely the model. It is the loop with no edge, the handoff that drops what mattered, the fan-out that multiplies exposure, and the checkpoint nobody wired in before an irreversible action.
A single successful run shows a system can succeed once. Pass@k, pass^k, and repeated-trial statistics are how anyone finds out whether it succeeds reliably — and task length changes the answer.
A working agent is a loop with a contract — observe, plan, act, reflect — plus explicit rules for what persists, what a tool call returns, and where a human has to say yes.
Every era of AI agent architecture answers one question differently: what decides the next action. A 1971 planner, a learned reward signal, a prompted reasoning loop, a delegated sub-agent, and a shipped SDK are five dated, sourced answers.
A single loop, a plan handed to an executor, a supervisor delegating to workers, or a graph of nodes and gated edges — four documented wiring patterns for the same underlying control problem, each with a stated price.
An agent architecture is five engineering decisions wired into one harness — permission scoping, state persistence, checkpoint gates, retries, and instrumentation — each documented by a real SDK, each with a demonstrated production failure mode.
Two questions decide how agent architecture matures as an engineering discipline by 2035: pattern convergence or fragmentation, and certified reliability or empirical best-effort. Four scenarios, each with a falsifier.
Claude Code ships its own permission system, hooks, subagents, and memory files. Each is documented, each has already failed in a specific and traceable way, and each failure is distinct from a generic "AI agent risk."