Agentic AI

The essays collected here treat agentic AI less as a mysterious property of a model than as a new software form — one whose constraints, permissions, and feedback loops are as decisive as the underlying model’s intelligence.

Claude Cowork makes the shift visible because the system does not merely answer prompts; it plans, executes, and loops through work. The essay opens with a frame that has stuck — analysts calling it the “SaaSpocalypse,” a watershed moment “where general-purpose AI agents began to demonstrably dismantle the business models of entrenched software-as-a-service (SaaS) providers.” The product story is itself revealing: Cowork was built in roughly ten days by four engineers, “using Claude Code itself to generate most of the code.” That recursive origin matters because it shows what changes once a system can do work rather than answer questions: “With chatbots, you ask a question and receive a response. With Cowork, you assign a task — and it doesn’t ‘guess’ the answer. It executes a loop of planning and action.” A planner breaks the request into sub-tasks; an orchestrator spawns parallel sub-agents for things like sorting hundreds of receipts or summarizing fifty papers.

The Fork in the Road: Claude Code vs CoWork sharpens the same point by drawing a hard architectural line. Both products share DNA, “yet profound differences lie beneath the surface.” Claude Code “runs on your machine” — a CLI wrapper that delegates execution to your filesystem, your shell, your dependencies; “Claude itself never owns the environment — it only reasons about it.” Cowork “provisions a VM and gives the agent its own computer,” selling reproducibility, isolation, and persistence as features rather than nice-to-haves. The essay’s claim is that this single architectural decision — who owns the execution environment — “profoundly shapes what each system can ultimately become.” Resumability, idle cost, security boundaries, and the kinds of long-running tasks each can support all flow from that one choice.

Prompt, Context, Harness supplies the theoretical scaffolding. Its three-stage history — phrasing as the problem, then context, then the harness — culminates exactly where these agentic products sit. By 2026, “agents were no longer a concept in academic papers or a promise in product announcements — they were showing up in actual engineering pipelines. And they brought their own class of problems.” The essential question is no longer whether one model can produce a strong answer in one turn, but whether a system “can stay coherent across tools, sub-agents, permissions, and changing state.” The surgeon-anesthesiologist analogy is the essay’s clearest articulation of why a single agent cannot be the whole answer: many problems are not harder versions of one-agent work but structurally different.

Read together, these pieces argue that agents force software to be rethought around execution and coordination. The connection to voice technology is not incidental — once an agent can carry out work in the background, the natural front-end for many tasks becomes conversational rather than graphical, and the dream of a competent voice companion stops being science-fiction and starts being a product roadmap problem.