Sign in Subscribe

Topic

ai-architecture

A collection of 39 issues

The Event-Driven Operator: Agents That Wake on Events, Not Cron

Polling-based agents fail by construction — not occasionally, but structurally. The event-driven operator pattern eliminates that failure class by separating wake-up, execution, and recovery.

The Supervisor Pattern: How to Use Cheap Models Without Losing Control

Cheap models can expand execution, but the architecture must preserve scarce judgment. The Supervisor Pattern separates bounded work from consequential review.

Seven Models Judged Each Other. The Mistakes Were the Most Useful Part.

Seven AI models were given the same prompt in blind OpenCode sessions and asked to rank all seven, including themselves. The mistakes — silent model substitutions, hallucinated benchmark scores, drifting price denominators — turned out to be more interesting than the leaderboard.

When Validators Matter More Than Better LLMs

In a multi-step AI pipeline, the dangerous failure is not just a bad model output. It is a bad intermediate result becoming trusted state. Validators matter because they stop bad state at the boundary.

Temperature 0 Does Not Make LLM Output Deterministic

Temperature controls token selection, not the full inference path. Learn where LLM output variation enters and how to test the right reliability contract.

The Prompt Is a Hyperparameter

Manual prompt engineering is optimization without an objective function. DSPy reframes prompts as hyperparameters, programs as signatures, and optimization as a first-class operation — and in doing so, shifts the core skill from prompt craft to metric design.

Text-First Extraction Is an Epistemic Error

RAG pipelines degrade in a pattern consistent enough to be structural: a text extractor destroys semantic layout before a single token reaches the model. The failure is not in the retriever. It is at the extraction boundary — where two-dimensional structure gets linearized into noise.

Durable Execution: Why Python while Loops Are Not Production Ready

A while loop is a state machine whose state lives in RAM. One pod eviction and every completed step is gone. Production agentic workflows require exactly-once semantics — a guarantee that must be encoded in the architecture, not hoped for through retry logic.

The API Is the New Command Line: A Manifesto for Agentic Design

The command line gave humans a grammar for commanding machines. The API gives agents the same grammar — at scale, in parallel, without a human in the loop. Every design decision downstream flows from whether you've internalized that shift.

Dumb Core, Smart Edge: Agentic Design

Many agentic systems I've watched fail in production had the same shape: intelligence concentrated at the center, where it was hard to test, replace, or reason about. The orchestrator was doing too much — routing requests, holding domain knowledge, managing memory, selecting tools, and shaping outputs, all at once.