Cosine

Snapshot from our London panel with @OpenAI on deploying AI coding tools at enterprise scale. 🙌 The talk covered risk, security, and governance as AI moves from experimentation to real-world use. Big thanks to @ukhomeoffice, @SciTechgovuk, and @Babcockplc for the great talk.

90% of tech professionals use AI at work. That means most teams are merging code nobody fully understands. Arguing about whether models write better or worse code than humans is a waste of time. There’s a far more important conversation happening around trust... Software engineering was built around a simple assumption: the person who made the change understood the change. With flippant adoption of AI, that assumption is starting to fail. We now have a world where code can be produced very quickly, approved casually, and merged because nothing obvious broke. Moving at this kind of speed is dangerous if ownership isn’t tight. The issue with AI-generated code is not just that some of it is sloppy - human-written code is sloppy all the time. The problem is that the old quality systems of engineering were built for sloppy code that still had authorship behind it. Someone knew why it was there and could explain the tradeoff. Teams might feel faster at the point of generation, but they’re now slower where it's most important: in review, maintenance, and trust. The most impactful AI companies in software will be the ones that can operate inside real codebases with enough context, discipline, and reliability that teams can actually treat the output as engineering work rather than autocomplete. At @CosineAI we recognise that the true challenge lies in making AI useful in the hard parts of software development: → Comprehending an existing system → Determining the correct change → Ensuring that the change is merged reliably We address this directly by requiring human review for every merge and providing transparent AI usage analytics for audibility and trust. I think that is where this whole market is heading. Less fascination with the generation, and much more scrutiny on ownership, comprehension, and whether the work actually holds up inside a living codebase.

A year ago Cosine was a much smaller team. We’ve now doubled in size, and we’re already running out of desk space. London feels like a home for AI right now. The community is fast, the talent is real, and it’s great to be building here. In 2 weeks we take the next step. 👀 Also, I'm sure every member of the team can confirm that putting together those desks is harder than it looks.

The recent AI-related incident at Amazon (6.3 million lost orders) should make the whole industry pause and ask an important question: Can we trust AI in real production environments? The short answer is… it depends. Let me explain why. After a string of engineering incidents, including one tied to its AI coding assistant and another that led to 6.3 million lost orders, Amazon is now tightening controls on how code gets written, reviewed, and shipped. That tells you everything about where the market is. We rushed to celebrate AI coding speed before we built the discipline to trust what was being produced. The AI coding market obsessed over autocomplete demos and ignored the harder question of production trust – where security, auditability, and uptime actually matter. That gap is now showing up publicly, and the consequences are significant. Developers are already telling us what is wrong. @StackOverflow found that 66% are frustrated by AI outputs that are “almost right, but not quite,” and 45.2% say debugging AI-generated code takes more time. This lines up with conversations I’m having with our customers: general AI tools often shift the work down the line rather than genuinely automate it. It’s clear that companies don’t need agents that generate more code, but systems that can generate code they can actually trust in production. Agents must be secure by design, fully auditable, and collaborative by default, with real accountability for how decisions were made and how changes got shipped. At @CosineAI, we believe the next generation of coding agents requires a fundamentally different approach. For AI to successfully build production systems, it must first earn production-level trust. In the coming weeks, we will be launching our vision for this. Stay tuned for updates.

Full report: cosine.sh/blog/benchmark…

Cosine is launching its first evaluation of LLMs on 4 critical coding languages: COBOL, ABAP, Rust, and Verilog. The strongest models fail on 50%+ of these specialist tasks. Mainstream coding performance is not a reliable indicator of real-world competence. Report below.

Is the COBOL migration dead? You might think that’s extreme, but hear me out. For decades, the rewrite justification was simple: the COBOL talent pool was vanishing (65% retire by 2030). The default answer was massive, multi-year, high-risk migrations. AI kills that logic. If AI can close the knowledge gap – helping engineers understand code, generate tests, and create docs – why take the huge cost and risk of translating millions of lines of working code? Selective modernisation, enabled by AI, is the smarter strategy. Keep the core stable, make maintenance easy with AI, and only rewrite with a genuine business case. Maintenance > Full Rewrite. This is exactly what we're solving at @CosineAI. Want to make your legacy systems a competitive advantage, not a liability? Let's talk.

The marketing team is going to hate me for posting this. 10 minutes unedited. Our coding agent refactoring a real production codebase. A single prompt, 1,857 lines refactored. The engineer set the task and constraints, stepped away, then came back to review. If your agent still needs a chaperone, it’s not autonomous.

Stop calling your AI coding tool automation. You're probably just using an autocomplete, and it muddies the water for teams trying to adopt AI seriously. People talk about AI coding tools as if they all belong in the same category. They don’t. There’s a real spectrum here, and most teams still struggle to map it. At one end, you have autocomplete. It’s useful, fast, and familiar. But the human is still doing almost all of the real work. They still have to manage context, decide what matters, and stitch the solution together piece by piece. In the middle, you get agentic assist. Now the model is not just suggesting the next line, it is helping edit, review, refactor, and reason across a larger surface area. This is where a lot of products stop, and where a lot of teams mistakenly think they’ve reached true automation. They haven’t. Task automation is a different operating model entirely. The system is not helping you type, but taking ownership of outcomes. Agents can pick up a ticket, execute against it, recover from failure modes, and bring back something reviewable. This is a different relationship entirely between humans and software. That shift is easy to describe, but harder to internalize. Many teams are still evaluating these systems with the wrong rubric. They ask whether the model is better at producing code, copy, or outputs in isolation. A more useful test is whether it can reliably handle the messy middle: incomplete requirements, fragmented context, changing constraints, and the inevitable ambiguity that exists in real production. That is where the difference between assistance and automation becomes obvious. In practice, this means the value of AI is no longer just in generation. It is in orchestration, judgment, and the ability to carry context through to a result that a human can review and steer. The human does not become irrelevant. But their role increasingly looks less like manual production and more like prioritization, oversight, and decision-making. I believe this transition is bigger than many people realize. We are not just improving the interface for getting work done, but also redefining who does which parts of the work in the first place. @CosineAI is building a software engineering agent that can take ownership of real engineering work in existing codebases, carry messy context through ambiguity, and return reviewable outcomes. See the difference for yourself.

If your commitment is to quality and durable enterprise results, we want to partner with you. 🔗Explore and connect: cosine.sh/partner

Our Partnership Program is officially live! We're building a network to accelerate autonomous software engineering in the world’s most critical and regulated enterprise environments. Our focus: security, sovereignty, governance, and tough legacy stacks like COBOL + Fortran.

Say hello to Sarah, the newest member of the Cosine team, as our Head of Marketing! With a strong background in B2B SaaS and enterprise tech, she’s leading our marketing strategy – shaping our narrative, building the brand, and helping us scale rapidly. Let’s go! 🚀

Nothing beats hearing how Cosine helps developers ship faster, stay in flow longer, and get more done. 🔥 We’re helping engineers build more with less friction. Your feedback shapes what we build next.

We're on to you 👀

"Copilot" is the wrong mental model for where AI development is going. A copilot sits next to you. Suggests things. Waits for you to fly. An autonomous engineer takes a ticket and comes back with a PR. They're not the same product category. We treat them like they are.

@dimitrioskonst Wait a minute...

I'm about to get fired. Joined Cosine (YC W23) 10 days ago. This weekend I rebuilt the entire website without telling anyone. New positioning. New copy. Built for developers. My team finds out when you do (for real)

Our prediction for the next 12 months? An enterprise AI backlash. The casual rollout we’re seeing across industries is creating a massive gap between policy, regulation, and reality. Here are 3 rules for secure rollouts in production:

I’m tired of seeing 10x productivity demos from off-the-shelf AI models that would cause major problems in actual production codebases. Those at the adoption stage need to invest in AI strategically and with caution. Here are 5 risks to watch for: 🧵

They cry out and demand for more we can not satiate the demand for more CLIs

We're thrilled to welcome Dimitrios, Cosine's new Head of Growth! 👋 @dimitrioskonst will lead our growth initiatives, with a focus on making it effortless for engineering teams to adopt our autonomous coding agent via self-serve. Welcome to the team!