Tagged agent architecture · 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.
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 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.
A language model has one input channel and no way to mark part of it inert. Everything it reads is a candidate instruction, which moves the defensible boundary off the prompt entirely and onto the authority behind every action the system is allowed to take.
Codex is best understood as a changing stack of models, tools, execution boundaries, state, and review mechanisms—not as a single code-generating model.