8th Light

Just when we thought automation could cover it all, the truth hits hard: you can’t replace the human experience! In the world of accessibility, automated tools like scanners and LLMs give us beneficial insights, but they still miss a lot. Research shows: - Traditional scans can miss 50-70% of accessibility issues. - LLMs often detect even fewer violations than those same tools. - False positives can lead teams to chase unnecessary fixes, wasting precious time. What truly matters is a human touch. A specialist with lived experience understands the nuances and frustrations that automated tools can't. They tap into real user flows and test with actual screen readers, tasks no algorithm can replicate. At 8th Light, we're committed to bridging this gap. Our experts ensure your digital experiences work for everyone. Don’t let automation be your only strategy. Real accessibility requires a genuine connection! Let’s start a conversation about making your digital world welcoming for all! #AccessibilityMatters #HumanExperience

481% increase in delivery during a platform rebuild. Stephen Walker built the system that made it possible: Tightrope. My fellow 8th Light engineer Stephen Walker helped a legal tech company achieve a 4x increase in pull requests while rebuilding their commerce platform and preparing for a high-volume launch. The challenge: Rebuild their product catalog, decouple legacy systems, migrate from .NET to Java, and adopt AI-assisted development - all before a mid-year launch handling tens of millions in transactions. The results: 481% increase in code shipped since January, stable 5% rollout achieved ahead of schedule, and a mid-year launch on track. Stephen and the team built Tightrope: a harness engineering workflow built on multi-agent orchestration. An orchestrator delegates to specialized sub-agents: planners, architectural reviewers, requirements reviewers, builders, and testers. Each agent operates in iterative loops until consensus is reached. After each PR, a retro updates the knowledge base so learnings compound across future work. Key design decisions: - Strict guardrails: every mutation required an explicit script call, preventing uncontrolled changes - Determinism hierarchy: scripts and hooks are most predictable, followed by skills and rules, then Claude.md files and agent memory - Managing agent context: condensed documentation without losing information, plus sub-agent orchestration to prevent context rot - Self-learning loops: insights from one PR guided future work The outcome: Engineers moved from writing code to evaluating it. They spent 30-40% of their time in review cycles, yet the PRs produced were high quality and far more numerous than before. The critical shift: Tightrope's self-improvement cycle shares knowledge across the entire team through the repo's agent context files. This moves beyond individual developer augmentation to collective intelligence improvement at the project level. Engineers who had little AI-assisted tooling experience were suddenly operating at a different level. What's your biggest blocker to scaling delivery while redesigning systems? #AI #SoftwareEngineering #DeveloperProductivity

Want to learn what's really going on when pairing software engineering with AI? In our new webinar series, AI in the Open, Travis Frisinger and Eric Olson dive into engineering workflows that bring software craft to working with LLMs. We’ll explore harness engineering, structured RPI loops, and multi-agent TDD patterns that keep humans in the loop and transform AI from a code generator into a reliable engineering partner. Registration below!

I set up CLAUDE.md with (what I thought was) a very explicit rule: never commit directly to main. Claude acknowledged the instructions and committed directly to main anyway. I asked why. Claude said you're right, I shouldn't have done that — and then did it again and again. Clearly, more explicit instructions weren't going to fix this. That's when I reached for hooks — a way to inject a check at the actual moment of action rather than just at session start. First, I tried a "warn" hook that fires before every file edit. Better than CLAUDE.md alone, but Claude saw the warning and proceeded anyway. That's how warn hooks are designed — they inform, but they don't stop. What do you do when Claude independently decides to ignore its warnings? 🙃 So I upgraded to a "block" hook — a shell script that checks the current branch and blocks the edit if it's main. Claude hit it, created a branch, and moved on. Finally, something it couldn't override. There's a difference between a rule Claude has read and a mechanism it can't bypass. I don't know if I fully appreciated that until I built both. A warning it ignores is not a guardrail; it's just documentation. Where in your workflow are you relying on Claude reading the instructions — and where have you wired in something it can't route around?

With harness engineering, system design is the new bottleneck. Can your systems produce reliable output at scale? Harness engineering is the discipline of designing systems that make autonomous output reliable, aligned, and repeatable. It covers how work is specified, how constraints are enforced, how context is made accessible, and how outputs are evaluated and corrected. When those systems are weak, even strong models produce inconsistent results. When they are well-designed, output becomes scalable and predictable, shifting the constraint away from production capacity and toward system design. This starts to look less like programming and more like managing a system of autonomous contributors. Some early signals point to the scale: OpenAI has described building production systems with large volumes of agent-generated code in a fraction of typical timelines, and Stripe has reported internal agents generating over 1,000 pull requests per week, written end-to-end by agents and then reviewed by humans. The pattern is consistent: output scales quickly, reliability does not. At 8th Light, we’re starting to see clients move in this direction and beginning to deploy these systems with them. The gap between teams with strong harnesses and those without is already noticeable. Where this shows up: 1) The bottleneck shifts to clarity If goals, constraints, and success criteria are ambiguous, output quality degrades quickly because the system can only operate within what has been explicitly defined. 2) Undocumented knowledge becomes unusable Anything not captured in accessible, structured form is effectively absent. Documentation becomes critical rather than nice-to-have. 3) Differences in operating model compound Two organizations using the same models will diverge based on how they structure work, capture decisions, and enforce feedback. 4) Governance is more tightly coupled to production Quality cannot rely on downstream review alone. It has to be encoded initially in constraints, then in tests and continuous monitoring. 5) Technical work shifts toward system definition Less effort goes into producing code directly, and more goes into defining intent, boundaries, and evaluation so systems can operate independently. Harness engineering is not a tooling upgrade. It is a shift in how work is defined and controlled around agentic systems. AI increases production capacity, while harness engineering determines reliability and outcomes. Organizations that treat system design as a first-class capability will see more consistent results. #agentic #ai #engineering

I set up CLAUDE.md with (what I thought was) a very explicit rule: never commit directly to main. Claude acknowledged the instructions and committed directly to main anyway. I asked why. Claude said you're right, I shouldn't have done that — and then did it again and again. Clearly, more explicit instructions weren't going to fix this. That's when I reached for hooks — a way to inject a check at the actual moment of action rather than just at session start. First, I tried a "warn" hook that fires before every file edit. Better than CLAUDE.md alone, but Claude saw the warning and proceeded anyway. That's how warn hooks are designed — they inform, but they don't stop. What do you do when Claude independently decides to ignore its warnings? 🙃 So I upgraded to a "block" hook — a shell script that checks the current branch and blocks the edit if it's main. Claude hit it, created a branch, and moved on. Finally, something it couldn't override. There's a difference between a rule Claude has read and a mechanism it can't bypass. I don't know if I fully appreciated that until I built both. A warning it ignores is not a guardrail; it's just documentation. Where in your workflow are you relying on Claude reading the instructions — and where have you wired in something it can't route around?

Congratulations to 8th Light's Spring 2026 Promotion Recipients ✳ Jessica Lupanow, Jonathan Manahan, Taylor Keazirian, Rani Zilpelwar, Scott Plunkett, Tony Distinti, Michelle Markus

Congratulations to 8th Light's latest Beacon Award Recipients: Kelsey Roy, Jonathan Manahan, and Eric Olson! The 8th Light Beacon Award is an honor that recognizes distinguished individuals or teams who bring 8th Light’s values to the frontlines, inspiring and enabling the way for others to follow. https://lnkd.in/eSqGasQe

A year or two ago, most companies were still experimenting with AI in limited, superficial ways. A chatbot here, a 'smart' recommendation engine there. Nothing mission-critical. We are far beyond that now. Behold the era of Agentic AI, which really ramped up alongside LLM maturity in back half of 2025. Today, AI has become embedded in operations, helping brands expand and experiment faster than ever. It’s accelerating breakthroughs in pharma, booking and rebooking travel, and transforming how we shop. In this issue of The Spark Engine, 8th Light’s newsletter, we explore how teams are using AI to think bigger while staying reliable and accountable to everyone, from medical patients to regulatory bodies. #thesparkengine