TRM's AI-Assisted Coding Adoption Model Boosts Productivity 125%

Late last year, AI-assisted coding adoption inside TRM's engineering org hit an inflection point. My instinct wasn't to roll out more tooling — it was to step back and rethink how my teams actually operate. I just published a piece on the TRM Tech Blog walking through the model we landed on: an "agentic software factory" that treats AI as shared platform infrastructure, not a collection of personal productivity hacks. The core idea is being explicit about human decision boundaries at every level. Some work is AI-first. Some is AI-assisted. Some is human-led, with AI organizing the evidence. And some — promotions, hiring, architecture tradeoffs, accountability calls — stays human-only, full stop. A few Q1 2026 numbers from running this on my blockchain data engineering teams: - 125% of OKRs completed (after adjusting for AI assisted delivery) - Chains onboarded at 3x our previous quarterly record - 35%+ reduction in targeted infrastructure spend - Zero confirmed P0 incidents tied to chain-launch scaffolding since rollout The thing I keep coming back to: when AI makes code generation cheap, the bottleneck doesn't disappear — it moves. To review quality. To prioritization. To engineering judgment. You can't just accelerate the old system; you have to redesign around the new constraint. If you're leading or building an engineering team right now, I think this is worth a read. Link in the comments. Always happy to compare notes.

A year ago, I told a few peers that developers wouldn’t be replaced but their roles would shift toward code review and oversight. Today, that prediction is playing out across the industry. AI coding tools are transforming engineering work from writing code to reviewing, debugging, and managing AI‑generated output. According to a recent report, 81% of engineering leaders say the time saved by AI is now spent reviewing its work, creating a growing layer of “invisible work” that traditional metrics fail to measure. This shift isn’t about losing jobs rather it’s about redefining them. The future of engineering will reward those who excel at AI‑augmented workflows, technical judgment, and high quality engineering. https://lnkd.in/g8xmkM74

“AI didn’t replace engineering. It changed where engineering happens.” There’s a lot of talk right now about vibe coding - using AI to rapidly generate software through prompts, iteration, and intent driven development. A lot of the talk is negative. And for good reason. Done poorly, vibe coding can create technical debt at machine speed: - Fragile architecture - Security gaps - Hallucinated dependencies - Code nobody truly understands - “It works” mistaken for “It scales” But done responsibly? It can be a force multiplier. I think of Responsible Vibe Coding as this: - AI accelerates creation. - Humans own judgment. That means we use AI to move faster on scaffolding, prototyping, repetitive logic, and exploration. We keep humans accountable for architecture, security, testing, scalability, and maintainability. We treat prompts as the start of engineering - not the replacement for it. We optimize for understanding, not just output. And we move fast, but don’t automate bad decisions The real shift isn’t that developers are writing less code. It’s that strong engineers are spending more time on system thinking and design decisions. The future probably isn’t AI writes software. It’s AI accelerated builders who know when to trust speed and when to slow down for discipline. That feels a lot more sustainable than chasing vibes alone.

The "Free Lunch" in AI Engineering just ended. Here is the bill. For the last two years, we’ve been living in a "subsidized" reality. We paid a flat fee and treated frontier models like they were free utilities. That era officially ends on June 1st. GitHub Copilot’s shift to usage-based billing isn't just a pricing update—it’s a market correction. The new "Model Multipliers" are a wake-on call for every Engineering Manager: Claude 3.5 Sonnet: 9x multiplier. Claude 3 Opus: A staggering 27x multiplier. If you're a Senior Engineer today, your value isn't just in writing code—it's in Orchestration. 1. Predictability over Novelty: A senior knows that a 27x cost spike for a task that could have been handled by a 1x model isn't progress; it's a failure of architectural governance. 2. Building the Harness: Real production systems (like Claude Code) are 98% deterministic engineering. It’s the context compaction, the recovery logic, and the permission gates that make the AI work. 3. The Economics of Code: We are entering an era where Model Selection is a core technical skill. Every token has a price tag, and every architectural decision has a multiplier. The future of Senior Engineering isn't about being replaced by AI. It’s about building the Systems of Governance that make AI safe and affordable for production. Are you architecting for the 27x reality?

AI is transforming engineering roles as it takes over some coding responsibilities, shifting job functions towards managing AI outputs, according to The State of Engineering Excellence 2026 report released Wednesday. Conducted by software platform Harness, the report surveyed 700 developers and engineering professionals at enterprise companies. The adoption of AI in engineering workflows has become standard, yet teams are struggling to measure its productivity impact and ROI. More than half of the respondents expressed concerns about performance evaluations based on AI data, emphasizing the need for a clear distinction between improvement data and their performance evaluations. Trevor Stuart, SVP of product and general manager at Harness, noted that AI coding is reshaping modern software development in unprecedented ways. He stated, “Engineering leaders are being asked to make multi-year AI investment decisions using dashboards built for a different era of software development.”

Everyone is talking about “Vibe Coding” these days 😂 But do you think building projects with AI is only about generating code fast? I don’t think so. The real skill now is: → Giving exact prompts with proper context → Making AI understand the complete architecture → Reducing unnecessary token usage → Getting accurate outputs in fewer iterations → Managing context efficiently in large projects Because in real companies, token consumption = cost. Soon, developers won’t just be judged by: “Who can code faster?” It may become: “Who can build production-ready systems with minimum AI cost and better prompting?” 👀 Prompt engineering is slowly becoming an actual engineering skill. Vibe coding is easy. Efficient AI-assisted development is the real game. What do you think — future developers or future prompt architects? 🚀

Another "food for thought bytes" ... AI-assisted coding is accelerating development, but it also introduces a new risk: trusting code that hasn’t been deeply understood. Just because generated code works correctly today doesn’t mean it won’t create security or maintenance problems tomorrow. A real life scenario... A developer merges a Copilot-generated PR. The code passes CI/CD. Peer review looks good. No obvious issues. Two weeks later, a security vulnerability surfaces. The security team flags it. Now everyone is asking: * Where did this code come from? * Who approved it? * Was it reviewed properly? * Is the same pattern repeated elsewhere in the codebase? And that’s the bigger issue. AI-assisted development is accelerating software delivery, but many organizations still lack: * AI code governance * Traceability * Security review standards * Provenance tracking * Clear accountability The risk isn’t that AI writes bad code. The risk is that developers trust code they DID NOT FULLY reason about. “Generated by AI” should never mean “assumed to be safe.” AI is a powerful accelerator — but responsibility still belongs to engineering teams.

I keep reading stories about engineering teams burning through massive AI token budgets. Some of it sounds almost funny at first. Developers bragging about how many tokens they used. Agents running for hours. Huge context windows. Massive refactors. “Look how much work the AI did.” But this is going to become a serious engineering management topic very quickly. If every project starts operating like this, AI spend will not be a cute line item anymore. It will become a meaningful part of engineering cost. And at that point, token efficiency becomes part of engineering performance. Not in a silly way where we punish people for using AI. The goal is not to use fewer tokens. The goal is to create more value per token. That means teams will need to measure things like Cost per merged PR Cost per resolved ticket Cost per production feature Cost per incident avoided Token spend per successful task How much agent work was actually useful How often the model loops, rewrites, or explores blindly The engineers who win will not just be the ones who “use AI the most.” They will be the ones who know how to use AI with discipline. Better prompts. Better context. Smaller scopes. Cleaner repos. Reusable instructions. Good architecture. Good tests where they matter. Clear task boundaries. Less blind agent wandering. The irony is that AI makes fundamentals more important, not less. Messy systems will burn tokens. Clear systems will compound. I think token efficiency will become one of the next serious engineering metrics. Not because finance wants to ruin the fun. But because waste at scale eventually becomes strategy’s problem.

Everyone wants to “build with AI” until they realize AI is the easy part. This week, I had an insightful conversation with a CTO who wanted to see what AI-powered software development actually looks like beyond the demos and hype. What stood out after the session was this: Subscribing to AI tools is easy. Building the engineering discipline, guardrails, workflows, and review systems to make AI work reliably inside a real development lifecycle is the hard part. The real craftsmanship is not in generating code. It’s in knowing: - what should be automated, - what needs Human-in-the-Loop validation, - where reliability matters more than speed, - and how to balance “ship fast” with “ship stable.” AI-assisted engineering is not vibe coding at enterprise scale. The teams seeing real outcomes are the ones treating AI like a force multiplier for strong engineering practices not a replacement for them.

Everyone is “vibe coding” now. But almost nobody is talking about the downside. AI can generate 1,000 lines of code in minutes. But can it understand: - system architecture? - scale? - security? - business tradeoffs? - production failures at 2 AM? That’s where the real engineering still starts. Right now, I see 3 kinds of developers emerging: 1. Developers who ignore AI → becoming slower every month. 2. Developers who blindly copy AI code → creating future technical debt bombs. 3. Developers who deeply understand engineering + use AI as an accelerator → these people are becoming insanely productive. The dangerous part about AI coding is not bad syntax. It’s false confidence. AI will confidently generate: - vulnerable code - fake APIs - inefficient queries - broken architecture - insecure auth flows …and many developers ship it without fully understanding it. My opinion: AI should probably write: - boilerplate - repetitive logic - documentation - CRUD APIs - test cases But humans still need to own: - architecture - scalability - performance - security - engineering judgment The best engineers in the next 5 years won’t be: “people who code everything manually” or “people who let AI do everything” It will be engineers who know exactly: WHEN to trust AI and WHEN not to. AI is not replacing software engineers. It is replacing engineers who stop thinking.