{"id":171141,"date":"2026-09-01T12:27:19","date_gmt":"2026-09-01T11:27:19","guid":{"rendered":"https:\/\/www.intelligentcio.com\/eu\/?p=171141"},"modified":"2026-09-01T12:27:20","modified_gmt":"2026-09-01T11:27:20","slug":"maintaining-software-integrity-as-ai-accelerates-development","status":"publish","type":"post","link":"https:\/\/www.intelligentcio.com\/eu\/2026\/09\/01\/maintaining-software-integrity-as-ai-accelerates-development\/","title":{"rendered":"Maintaining software integrity as AI accelerates development"},"content":{"rendered":"\n

Tom Totenberg, Head of Release Automation and Observability, LaunchDarkly, examines how organisations can maintain software quality, confidence and structural integrity as AI accelerates code creation and software delivery.<\/em><\/p>\n\n\n\n

Away from screens, I\u2019ve occasionally dabbled in making things out of wood. Poorly, mind you, but the process and craft is something I find fascinating.<\/p>\n\n\n\n

There are plenty of parallels between the principles of woodworking and the rigour of the traditional software development process. Take the well-worn adage of \u2018measure twice, cut once\u2019. Or, in my case: measure six times, cut once, realise I still somehow messed up, measure a few more times then cut again. It\u2019s a feedback loop of thinking, doing and dry-fitting. It\u2019s meditative. It\u2019s nice!<\/p>\n\n\n\n

Software engineering used to afford us similar moments of pause. We had the luxury of inspecting puzzles, thinking deeply and having stimulating human-to-human conversations that acted as our \u2018dry-fit\u2019, ensuring every line of code sat flush with the existing architecture.<\/p>\n\n\n\n

AI, however, and all the expectations that it brings, has rapidly stripped this away. The technology promised to make software development faster and it certainly has delivered. Inconsistently and not everywhere, but it has delivered.<\/p>\n\n\n\n

What was once a craft practised at the workbench is starting to look more like a software factory, with change moving continuously rather than one carefully handled release at a time. Code creation velocity has increased, compressing timelines and shifting bottlenecks elsewhere.<\/p>\n\n\n\n

The danger of working towards tighter timelines is that they force even the most diligent engineers to rely on connectors that weren\u2019t intended to be load-bearing but suddenly become critical infrastructure. (We\u2019ve all seen that XKCD comic. If not, you\u2019re one of today\u2019s lucky 10,000!) Many take shortcuts by removing the peer review process, instead opting to ask a large language model to generate a complex workflow.<\/p>\n\n\n\n

Our rubber duck talks back! This is incredible! It sounds confident and I\u2019m mentally fatigued enough to believe it!<\/p>\n\n\n\n

This weakens the traditional safety nets and as production accelerates further, so too do the number of emergency incident meetings and lengthy recovery times. When output starts moving at machine speed, humans inevitably can\u2019t keep up.<\/p>\n\n\n\n

Our own research at LaunchDarkly highlights the scale of this problem. We found that while AI has increased code velocity for 94% of engineering and DevOps professionals, it decreased confidence for 91%. Over eight in ten (81%) have shipped code with unresolved risks within the last six months alone, creating a cycle where 38% spend more than a quarter of their time resolving incidents.<\/p>\n\n\n\n

The practice is still emerging, but almost every technology leader I speak with is thinking about ways to automate their SDLC with AI agents. We\u2019re collectively still learning and inventing how to do this, but the allure of a software factory is undeniable: measure twice, cut indefinitely.<\/p>\n\n\n\n

Closing the loop between monitoring and production impact<\/strong><\/p>\n\n\n\n

Human engineers becoming more distant from individual lines of code has distinct ramifications on remediation times. The irony is that our observability stack has never been more sophisticated: unified telemetry, AI-powered dashboards and real-time anomaly detection that can alert DevOps engineers when a system starts to crack.<\/p>\n\n\n\n

Yet when something breaks in production, teams still ask the same questions: which change caused this? Who has the knowledge to fix it?<\/p>\n\n\n\n

This is how we end up in everyone\u2019s favourite 2:00 AM war rooms. Wake everyone up, scramble to diagnose the issue and hope you can identify it fast enough that some teams get to go back to sleep.<\/p>\n\n\n\n

The lack of feature-level observability has been undermining the efficiency and accuracy of incident response for years. More signals cannot fix the problem if the platform can\u2019t correlate issues to the changes that exposed them.<\/p>\n\n\n\n

Until recently, solving this was a big tech luxury, with the capability reserved for those with deep pockets, huge engineering teams and bespoke internal tools. As AI brings big factory-level velocity to every organisation, having a system that connects releases directly to production impact becomes a necessity, not a luxury.<\/p>\n\n\n\n

Who knows, maybe there will still be a market for artisanal, hand-crafted, farm-to-table software. Head to the market, buy some fresh eggs and pick up a carefully created application from someone in overalls with a laptop.<\/p>\n\n\n\n

For the rest of us poor schmucks, we must create the infrastructure for safe, repeatable and automated software delivery.<\/p>\n\n\n\n

Paving the way for sustainable implementation<\/strong><\/p>\n\n\n\n

To reclaim confidence, we need pipelines supported by the right tools, techniques and infrastructure. With modern AI\u2019s capabilities, the potential for automation explodes. However, it\u2019s not magic and we need to wield it responsibly.<\/p>\n\n\n\n

One useful distinction is between AI agents and deterministic code. Agents are flexible, but they\u2019re also expensive, slow and difficult to trust. Deterministic code, by comparison, is rigid, but it\u2019s cheap, fast and, with good testing, trustworthy.<\/p>\n\n\n\n

As with woodworking and pre-AI software, fast feedback loops in your software factory are the best way to ensure quality. Agents are an exemplar: their flexibility is both their biggest benefit and sharpest drawback. They require constant monitoring.<\/p>\n\n\n\n

So, as you start automating your pipelines with agents, here are a few concepts which work well for us at LaunchDarkly:<\/p>\n\n\n\n

Don\u2019t trust your agents.<\/strong> Every individual agent needs hard capability boundaries, isolated workspaces and zero trust permissions. (Read the entertaining Anthropic blog example here, where an agent without permission for a task slacked another agent to do something for it.)<\/p>\n\n\n\n

Seriously, don\u2019t trust your agents.<\/strong> After every agentic step, surround it with immediate verification. Two main flavours here: deterministic verification scripts to test whether the agent\u2019s claims are true and LLM-as-judge to score the agent\u2019s output.<\/p>\n\n\n\n

Prioritise production validation.<\/strong> Feature-level observability is the critical last mile of automation. This isn\u2019t optional; it\u2019s the only way that your production remediation and optimisation efforts can keep up with software factory scale.<\/p>\n\n\n\n

Maintain an orchestration layer.<\/strong> Pipelines are rarely linear and your teams\u2019 overlapping work certainly isn\u2019t. Supervisory agents are the best way to resolve minor agent conflicts and surface larger issues to human orchestrators.<\/p>\n\n\n\n

Keep humans in some loops.<\/strong> The whole point here is to limit where human attention is needed, so that manual inspection is impactful and not overwhelming. Route human approvals based on service criticality and change risk level. Sample everything else. When you change the pipeline structure itself, inspect closely until it\u2019s trustworthy enough to back off.<\/p>\n\n\n\n

As code generation picks up speed, we cannot rely on the same manual checkpoints that functioned well in a pre-AI world. A true golden path utilises best-in-class solutions that serve as guardrails, ensuring developers can standardise configurations without worrying about production outages. These guardrails will not slow teams down, but instead provide the brakes that make high-speed releases safe.<\/p>\n\n\n\n

Ever driven a moving truck? Slow to start, slow to stop and with all your pieces of furniture knocking against each other inside. These are your traditional releases: all or nothing. The future is a swarm of autonomous drones each carrying an individual package. Quick, nimble, loosely coupled and able to turn on a dime.<\/p>\n\n\n\n

However, while our delivery must be independent, our observability must be cohesive. While there are many approaches to this, at LaunchDarkly we\u2019re big believers in OpenTelemetry.<\/p>\n\n\n\n

For our regression detection, which we call guarded rollouts, flag- and audience-enriched telemetry provides rapid high-confidence automated rollback. Humans and agents then have access to pre-filtered observability data, scoped to the regressed feature, to remediate the underlying issue. It\u2019s like easing into a newly built chair before you give it your full weight.<\/p>\n\n\n\n

Building confidence at AI speed<\/strong><\/p>\n\n\n\n

With careful instrumentation, validation gates and runtime control, we can ensure that AI-driven development remains a tailwind, rather than a liability. DevOps no longer needs to rely on best guesses about why a system is failing, instead using real-time diagnostics to keep it moving smoothly. This is key to moving from the 5% success rate of custom AI tools to a model of sustainable, high-velocity innovation.<\/p>\n\n\n\n

AI has changed the speed of the game, but it shouldn’t change the integrity of our output. The golden path leads to confidence. And if the future of software delivery is a factory, integrity can no longer depend on manual inspection at the very end. It has to be built into the line at every step. Even in an era of unprecedented velocity, the glue only sets on a release when we are certain the joints are square.<\/p>\n\n\n\n

OK, I\u2019ll stop with the woodworking analogies now.<\/p>\n","protected":false},"excerpt":{"rendered":"

Tom Totenberg, Head of Release Automation and Observability, LaunchDarkly, examines how organisations can maintain software quality, confidence and structural integrity as AI accelerates code creation and software delivery. Away from screens, I\u2019ve occasionally dabbled in making things out of wood. Poorly, mind you, but the process and craft is something I find fascinating. There are […]<\/p>\n","protected":false},"author":58,"featured_media":171142,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[21064,20673,29,93],"tags":[158,21655,577,1272,26613,22550,20667,16761,24431,22551,2537,23079],"class_list":["post-171141","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-expert-opinion","category-software","category-top-stories","tag-ai","tag-ai-agents","tag-artificial-intelligence","tag-devops","tag-feature-level-observability","tag-launchdarkly","tag-observability","tag-opentelemetry","tag-release-automation","tag-software-delivery","tag-software-development","tag-software-engineering"],"acf":[],"publishpress_future_workflow_manual_trigger":{"enabledWorkflows":[]},"_links":{"self":[{"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/posts\/171141","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/users\/58"}],"replies":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/comments?post=171141"}],"version-history":[{"count":1,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/posts\/171141\/revisions"}],"predecessor-version":[{"id":171143,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/posts\/171141\/revisions\/171143"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/media\/171142"}],"wp:attachment":[{"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/media?parent=171141"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/categories?post=171141"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.intelligentcio.com\/eu\/wp-json\/wp\/v2\/tags?post=171141"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}