A New Iron Triangle

When you think about it, at its core, software engineering has always been a capital problem rather than an engineering problem. The challenge remains to build something that solves a great need but is affordable by the organisation and its customers.

When I was a young consultant (I’m not going to embarrass myself by telling you how long ago that was; check out my LinkedIn profile if you’re curious), I was taught that the iron triangle of Time- Cost- Scope was inviolable. You can keep to a strict budget but if you do, you’ll need to choose whether to spend longer or achieve less scope and might end up doing both. Or you can keep to a deadline, but you’ll need to trade scope and probably spend more to do so. Us expensive consultants were needed to advise our clients on the right optimisation of this iron triangle for their market.

Now that we’re following an agentic software development lifecycle (ADLC), it seems that this iron triangle is no more. Time is no longer a constraint as agents can build code in a blink of an eye. Cost is also greatly reduced because LLM subscriptions are still vastly cheaper than consultant day rates. And if building out a software solution doesn’t cost much time, nor money, to build out an idea, Scope also stops being a constraint. It’s cheap to add as much scope as you can think of.

However, it’s a mistake to think that following the ADLC means no more constraints. As software engineering patterns become settled, it’s becoming increasingly clear that we now have a new iron triangle to optimise for our context: Autonomy- Tokenomics- Intent. As before, you can’t optimise them all, and you can optimise one of these sides only if you slacken the other two. Let’s examine them one-by-one.

The autonomy horizon doesn’t mean how long you can YOLO a vibe coding session before your agent accidentally-on-purpose wipes out your entire repo. Rather it is how long your agent + harness can run and still produce accurate results safely. It’s a culmination of the capability of the LLM, the extent of guard rails, the effectiveness of context engineering. It’s also a culmination of the effectiveness of the AgentOps function, how conversant engineers are with their tools, how well the environments are maintained.

Doing this list of work comprehensively will cost more in effort from humans and token burn from agents. The second side of the triangle isn’t simply cost because tokenomics conveys the idea that not all tokens are created equally. Model routing is the idea that sometimes a simpler LLM can be more cost-effective for simpler problems, but not for more complex problems because of the volume of tokens needed to convey context effectively to simpler LLMs. Specialised LLMs can be both smaller and need fewer tokens. Tokenomics considers ideas like getting LLMs to talk like a caveman can be very token-efficient; however, we have to spend tokens inefficiently to counter new LLM malware attack vectors. The biggest trade-off in Tokenomics is that spending less on tokens results in humans spending more time in the loop.

Intent is the third side of our triangle. The new work is no longer about creating outcomes but defining the problem space, orchestrating multiple agentic systems, human interactions, taking legal and moral responsibility, overcoming the necessary frictions. This transition from output to intent is the defining management challenge of the decade. If we optimise for intent fidelity, we will need to flex autonomy horizon and spend more on tokens, to keep course-correcting the agents. The inverse isn’t necessarily true; if we don’t optimise for Intent, e.g., if we use the very effective pattern of building-to-completion multiple solutions then testing to see which is best, we will incur much more token spend also. This is induced demand at work: just as wider roads encourage more traffic, the ability to burn more tokens gives the opportunity to burn them more creatively.

For a while, the transition to ADLC seemed like an engineering problem, and our efforts were spent on figuring out how to exploit models effectively. We now need to think ahead and remember how to treat developing using ADLC as the capital problem it always was.

© Mind Rocket Services Ltd 2026. All rights reserved.

Avi Sinharay

CEng, MIET, MEng, MA (Cantab.)

Fractional CTO, Director/VP of Technology

Core Expertise

Technology Leadership, AI Native Dev, Operating Model Design, Engineering Culture

Domains

Health Tech, Media Tech