https://unboundforce.dev/tags/architecture/Recent content in Architecture on Unbound ForceHugoen-USCopyright (c) 2025-2026 Unbound ForceSun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/8-phase-pipeline/Sun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/8-phase-pipeline/<h2 id="the-two-phase-illusion">The Two-Phase Illusion</h2> <p>Every team building AI agent workflows discovers the same thing: letting an agent plan and execute in the same pass produces unreliable output. The planning step must be separate, with its output reviewed before implementation begins. OpenAI, Anthropic, and ThoughtWorks all arrived at this conclusion independently (Yanli Liu, &ldquo;Harness Engineering,&rdquo; <em>AI Advances</em>, Apr 2026).</p>https://unboundforce.dev/blog/convention-packs/Sun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/convention-packs/<h2 id="the-prose-standards-problem">The Prose Standards Problem</h2> <p>Every project that uses AI coding agents needs coding standards. The agent needs to know how to format code, name variables, handle errors, structure tests, and manage dependencies. Without explicit standards, the agent guesses — and its guesses are inconsistent across sessions, files, and projects.</p>https://unboundforce.dev/blog/build-to-delete/Sun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/build-to-delete/<h2 id="every-component-is-a-bet">Every Component Is a Bet</h2> <p>Every component in an AI agent harness exists because someone believed the model could not do something reliably enough on its own. A multi-agent review council exists because one agent cannot reliably self-review. A specification pipeline exists because agents cannot reliably plan and execute in the same pass. Convention packs exist because agents do not consistently follow coding standards without explicit rules.</p>https://unboundforce.dev/blog/ai-code-reviewer-write-access/Sun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/ai-code-reviewer-write-access/<h2 id="the-feedback-loop-that-eats-itself">The Feedback Loop That Eats Itself</h2> <p>When the same AI agent writes code and reviews it, the review is compromised. The agent has context about why it made each decision. It remembers the trade-offs it considered. It is, in the most literal sense, reviewing its own homework.</p>https://unboundforce.dev/blog/five-principles-ai-agent-harness/Sun, 03 May 2026 00:00:00 +0000https://unboundforce.dev/blog/five-principles-ai-agent-harness/<h2 id="the-pattern-nobody-expected">The Pattern Nobody Expected</h2> <p>Three independent teams — OpenAI, Anthropic, and ThoughtWorks — each spent months building AI agent harnesses. They started from different assumptions, used different architectures, and optimized for different goals. They arrived at the same five conclusions.</p>