Notes on agents, software, and work.
Systems / August 16, 2026 / 10 min read
Generative AI agent architecture: a control and recovery diagram
Generative AI agent architecture needs an execution path, failure-aware recovery, application permissions, and limits on retries and high-impact actions.
Systems / August 16, 2026 / 10 min read
Agentic workflows vs AI agents: choose where variability belongs
Choose a fixed workflow, bounded agent, or open agent loop by matching task variability to control, recovery, observability, and cost.
Systems / August 16, 2026 / 11 min read
AI agent orchestration framework: choose by coordination cost
The right AI agent orchestration framework depends on whether your workload needs deterministic rules, one agent, or coordinated specialists.
Systems / August 16, 2026 / 11 min read
Agent scaffolding: where task setup meets the agent harness
Agent scaffolding surrounds an LLM with tools and structure; a responsibility map separates task setup from state, approvals, evaluation, and recovery.
Systems / August 16, 2026 / 10 min read
Agent iteration: build loops with verification and stop conditions
Agent iteration is a bounded observe, act, and verify loop with records, budgets, stop conditions, retry limits, and handoff rules.
Systems / August 16, 2026 / 10 min read
AI for solopreneurs: build a stack you can control
The best AI stack for a solopreneur fits one workflow's value, authority, evidence, cost, and recovery needs before any tool is chosen.
Systems / August 16, 2026 / 10 min read
LangGraph vs OpenAI Agent SDK: choose by control surface, not feature count
LangGraph fits persistence-heavy, finely controlled workflows; OpenAI Agents SDK fits scoped assistants, delegation, and built-in tracing.
Systems / August 16, 2026 / 10 min read
MCP servers architecture: choosing local, remote, state, and trust boundaries
MCP server architecture choices turn on execution location, maintenance, authorization, session state, and context cost, not tool count.
Systems / August 16, 2026 / 10 min read
MCP servers vs agents: where each layer belongs
Agents handle ambiguous, multi-step work; MCP servers expose tools and data. One architecture can combine both, with direct APIs for predictable paths.