https://toddschiller.com/2026-08-12T00:00:00-04:00Human ✘ Artificial Intelligence2026-08-12T00:00:00-04:002026-08-12T00:00:00-04:00Todd Schillertag:toddschiller.com,2026-08-12:/blog/parallel-agents-diverge-converge.html

Running coding agents in parallel is closer to a team of different developers than one developer multitasking, so plan a deliberate converge phase after the diverge phase.

<p>Running coding agents in parallel is less like one developer juggling features and more like a team of different developers building at the same time independently.</p> <p>Even when the agents share the same system prompt and the same skills, they diverge fast. LLMs are stochastic, and the divergence compounds with every decision.</p> <p>Spec-Driven Development and Domain-Driven Design help. They give each agent a global view of what the others are doing and a shared, ubiquitous language. But in practice, compounding drift erodes coherence fast.</p> <p>The only way to keep your codebase sane is to plan deliberate converge phases.</p> <p>Here are my tips for diverge-converge:</p> <ul> <li>Put new features behind a feature flag so agents can iterate on the intermediate result cleanly</li> <li>Define protocols and interfaces first to contain where drift can happen</li> <li>Use static analysis and linters to bound architectural and stylistic drift</li> <li>Use a skill to update your specs and ubiquitous language based on the gaps the diverge phase exposed</li> <li>Use a skill to identify and prioritize the tech debt the diverge phase introduced</li> </ul> <p>What's your approach to building features in parallel with coding agents?</p> <!-- markdownlint-disable MD013 --> <figure> <img loading="lazy" decoding="async" src="https://toddschiller.com/assets/images/parallel-agents-diverge-converge/parallel-agents-diverge-converge-landscape-2400.png" alt="Whiteboard-style diagram titled Parallel agents drift, plan to converge. From a single Shared start node (same prompt, specs, skills), three colored paths labeled agent 1, agent 2, and agent 3 fan outward through a pink Diverge zone where drift compounds with every decision, reach their widest spread, then are pulled back together through a green Converge zone into one Coherent result node. A footer reads: converge with feature flags, interfaces first, linters, and a spec and debt refresh."> <figcaption>Parallel coding agents fan out during the diverge phase, then a deliberate converge phase pulls them back into a coherent result.</figcaption> </figure> <!-- markdownlint-enable MD013 --> 2026-07-07T00:00:00-04:002026-07-07T00:00:00-04:00Todd Schillertag:toddschiller.com,2026-07-07:/blog/loop-engineering-compounding.html

Loop engineering for AI agents should raise the floor on each iteration by creating tools and correcting instructions.

<p>Loop engineering for AI agents should be about compounding, not repetition.</p> <p>A weak work loop does the same thing every pass: trigger, execute, review.</p> <p>A strong work loop leaves the system better than it found it.</p> <p>Sure, many harnesses now automatically improve via agent memory. However, that's leaving your work to chance: memory creation is inconsistent, and false memories degrade performance.</p> <p>The better way is to edit the loop itself:</p> <ul> <li>Skills/Instructions: conventions, assumptions, examples, failure modes</li> <li>Tools: scripts, workflows, fixtures, test harnesses</li> </ul> <p>With each run, the loop becomes more capable, more reliable, and more cost-effective.</p> <p>What are your go-to strategies for compounding?</p> <!-- markdownlint-disable MD013 --> <img loading="lazy" decoding="async" src="https://toddschiller.com/assets/images/loop-engineering-compounding/loop-engineering-compounding-landscape-ai-whiteboard-2400.png" alt="Whiteboard-style diagram titled Agent loops should compound. The left panel shows a repeating routine that starts from the same place each time: Trigger, Execute, Review. The right panel shows a compounding routine where loop 1, loop 2, and loop 3 climb upward on a system-improves axis, with orange step-up arrows labeled Refine and gains labeled Task done, plus instruction, and plus tool."> <!-- markdownlint-enable MD013 --> 2026-06-05T00:00:00-04:002026-06-05T00:00:00-04:00Todd Schillertag:toddschiller.com,2026-06-05:/blog/agent-browser-shield-product-hunt.html

agent-browser-shield, a source-available defense layer for AI agents in the browser, is on Product Hunt today.

<p>agent-browser-shield is live on Product Hunt today.</p> <p>Your AI agent that lands on a fresh page is one prompt injection away from leaking credentials, one dark pattern away from buying the wrong thing, and one fake review away from a bad recommendation.</p> <p>agent-browser-shield sits between the browser and the agent and blocks all three. Source-available. Ships as a Chrome extension and an OpenClaw skill.</p> <p>If you've been following the build, today's the day to upvote:</p> <p>👉 <a href="https://producthunt.com/products/agent-browser-shield">https://producthunt.com/products/agent-browser-shield</a></p> 2026-06-03T00:00:00-04:002026-06-03T00:00:00-04:00Todd Schillertag:toddschiller.com,2026-06-03:/blog/agent-browser-shield-june-03-update.html

14 new rules shipped in agent-browser-shield, now installable from the Chrome Web Store, and using it for daily driving.

<p>Two days after the alpha announcement, agent-browser-shield has a Chrome Web Store listing and 14 new protection rules.</p> <h2>Install from the Chrome Web Store</h2> <p>The extension is live at <a href="https://chromewebstore.google.com/detail/agent-browser-shield/gnejacdioaelglahihpagpfjpddpnamd">chromewebstore.google.com/detail/agent-browser-shield</a>. One click instead of unpacked-from-source. The prebuilt ZIP and source-build paths stay for Browserbase and other runtimes that need an unpacked extension.</p> <h2>New rules: handling prompt injection and context pollution in invisible surfaces</h2> <p>A browser-use agent reads surfaces a sighted user never looks at. The new rules close them:</p> <ul> <li><code>&lt;noscript&gt;</code> blocks (never rendered with JS on, but agents walk them)</li> <li>Poisoned <code>&lt;meta&gt;</code> description and <code>&lt;title&gt;</code> (the compact &quot;what is this page&quot; answer many agents pull first)</li> <li>JSON-LD <code>&lt;script&gt;</code> blocks (cited as the &quot;trusted summary&quot; of a page)</li> <li><code>aria-label</code>, <code>alt</code>, <code>title</code>, <code>placeholder</code>, and SVG <code>&lt;title&gt;</code> / <code>&lt;desc&gt;</code> / <code>&lt;text&gt;</code> (a11y-tree carriers)</li> <li>Unicode tag characters, bidi overrides, and zero-width payloads</li> <li>Long base64 / hex / percent-encoded blobs (the &quot;decode this and follow it&quot; pattern)</li> </ul> <h2>New rules: trust laundering</h2> <p><code>link-spoof-annotate</code> flags Cyrillic homoglyphs and anchors whose visible text doesn't match the href apex. <code>disguised-ad-flag</code> collapses native advertorials (Sponsored / Paid Post) that share DOM shape with editorial. <code>trust-badge-annotate</code> and <code>schema-trust-sanitize</code> ship off by default while we assess their false-positive rates.</p> <h2>The daily-driver surprise</h2> <p>I've started running it on my own daily-driver browser, not just agent runs. There have been some funny quirks (e.g., flagging GitHub issue links with &quot;.md&quot; in the link text as suspicious and hiding the GitHub issue template modal). However, overall it's been a positive on my browsing experience. So, we'll be experimenting with making more annotations visible to humans and multi-modal LLMs.</p> <p>⭐ <a href="https://github.com/pixiebrix/agent-browser-shield">https://github.com/pixiebrix/agent-browser-shield</a></p> 2026-06-01T00:00:00-04:002026-06-01T00:00:00-04:00Todd Schillertag:toddschiller.com,2026-06-01:/blog/agent-browser-shield-alpha.html

Announcing the alpha of agent-browser-shield — a source-available defense layer (also available as an OpenClaw skill) that blocks prompt injection and dark patterns before they reach your agent.

<p>There are 6 billion internet users. With AI agents, we're quickly heading to 60 to 600 billion &quot;users&quot; of the web.</p> <p>How do we keep all those agents safe when they touch the browser?</p> <p>At PixieBrix, we've spent years protecting BPO contact centers from insider risk, fraud, and social engineering in the browser.</p> <p>Today, we're applying that defense to AI agents and making a free, source-available, browser extension available (on GitHub and ClawHub).</p> <p>Your AI agent that lands on a fresh page is one prompt injection away from leaking credentials, one dark pattern away from buying the wrong thing, and one fake review away from a bad recommendation.</p> <p>Agent Browser Shield sits between the browser and the agent. It blocks:</p> <ul> <li>Prompt injection: visible or hidden instructions in page content</li> <li>Dark patterns: manipulative UI designed to trick/coerce</li> <li>Context pollution: low-value context that impairs instruction following</li> </ul> <p>A useful side effect: stripping irrelevant content also cuts token burn.</p> <p>Come join us on our mission. File issues, send PRs, or just tell me what you hate/love!</p> <p>The threat surface for agentic browsing is evolving fast. Let's defend our AI assistants together!</p> <ul> <li>⭐ Star the GitHub repo: <a href="https://github.com/pixiebrix/agent-browser-shield">https://github.com/pixiebrix/agent-browser-shield</a></li> <li>🦞 Star the OpenClaw skill: <a href="https://clawhub.ai/pixiebrix/agent-browser-shield">https://clawhub.ai/pixiebrix/agent-browser-shield</a></li> </ul> <p><img src="/assets/images/agent-browser-shield-alpha/demo.webp" alt="Side-by-side demo: an unprotected agent vs. one shielded by agent-browser-shield" loading="lazy" decoding="async" /></p>