
The precepts of software engineering are white hot right now. Hotter than Mugatu’s Winter collection. Hot enough to melt your face. Every week brings new methods of building – software factories, skills, agentic loops and more – which is incredibly exciting. Our industry is exploring what it means to have models fabricate whole systems at wild pace. But what does this compression mean really? Like, not what you tell your boss to maintain job security, or whatever cynical platitude you have memorised for the Substack comments section; what does it mean on a fundamental level, and how can we ride the wave rather than drowning in it?
Let’s take an example feature build – adding multi-step approval to our CMS pages. 12 months ago this might have taken weeks to design, build, and test. Assuming we understand the problem well enough, and have time to build, we might create documents, tickets, or similar to capture the analysis – desired states, NFRs – basically a bunch of artefacts used to document the proposal. We would build, test, fix, then ship – maybe getting feedback after a few weeks.
But today, we can simply trade tokens and get this done in an afternoon. What’s changed?
The Pace of Work
Four components that govern how quickly work gets done include:
- Knowledge creation – how quickly can we find something out? If we poke reality and measure how much it pushes back we have created some knowledge. Customers see typos and policy changes as completely different changesets. The customer already knew, the discovery is ours.
- Learning – the transfer of useful understanding between people. Moving from unaware, to deep understanding is the challenge.
- Resources – what we have available to use. People, compute, attention, time with customers, and anything else we can marshall.
- Arrangement – how the resources are connected and who can use them. A vault full of cash is useless to someone without the key.
Cost runs through all four – every experiment, teaching moment, research session and time spent waiting incurs cost. AI changes the economics because it reduces some of these costs to an extreme degree.
What agentic engineering changes
Existing models of software engineering were derived from hard engineering. If you miscalculate how much your rebar can hold people will die, so it’s really important to double-check your numbers. This process means spending plenty of time making sure we understand every failure mode, desirable state, risk, and plenty of other dimensions before we begin to build – because mistakes are expensive. Nobody wants to re-do their foundations several times over because concrete is pretty fucking against being broken up once set.
But agentic engineering means we can create knowledge by building real product, not artefacts like documents. This closely mirrors how the film industry went from, ahem analog film to digital editing. Editors went from physically cutting and pasting reels together, to creating multiple versions in parallel – showing the director and just keeping what they wanted. The importance of getting it right before building dropped way down. Multiple version creation still had a cost, but it was way, way below that of screwing up your physical film.
Some decisions still deserve sustained investigation because the costs remain high, or the knowledge that would produce a good decision has yet to be created. There are also irreversible processes like cooking an egg. I can’t unfry it no matter how much I want to. I can’t capture the approver’s IP address for auditing after the fact. We should therefore consider the downside risk of some decisions through careful attention to the side effects of our software. We can model these as costs – if the egg-chobbler requests scrambling (of their egg, not themselves) the cost is everything associated with re-doing their meal. If the auditor who reviews infosec compliance isn’t happy with the lack of IP logs and fails the business on account of that, the cost is…er…one less acronym on that company’s website? Or something? (Reminder: read our infosec policies again.)
There is also a subtle return on time spent living with and within a system. SREs know this well – they’re shaped by pain.

My favourite question for anyone working on complex system is tell me where your bottlenecks are. Okay that’s a directive. But you get the idea. This sort of knowledge always starts as tacit – in the minds of those that are responsible for a system, and if you’re lucky it makes its way into the system itself – encoded in gates, rules, and other explicit ways. But we shouldn’t underestimate the value a team holds in their implicit understanding of a product.
There is a solid return from writing and maintaining software over years which helps develops this knowledge – and builds big juicy judgement muscles which make engineers better over time. An agent can explain a codebase, but it’s only as good as what has been made explicit. Both a customer conversation nobody recorded and an outage who’s cause hasn’t been documented can still matter years later.
The return depends on where the human attention goes. An engineer that owns the page approval workflow may pay dividends over the course of years, asking them to build every variation of it has a different return, and a shorted development cycle gives us more choice about where to invest in depth.
Do you even agenticaly-loop bro?
The awkward thing about making engineering ten times faster is that the next person in the chain still has the same Tuesday afternoon.
In software land, product can become the first bottleneck. An engineer can generate five plausible versions before anyone has judged the first. Each is cheap to build, but each demands attention. Exploring alternatives pays when comparing them settles a decision. Otherwise a team exchanges a queue of things to build for a queue of things to review.
Good news. We’ve all become attention-hopping chimps trained by our lord and saviour TikTok, so perhaps the attention thing isn’t much of a problem.
Marketing has to make sense of what’s changing and generate enough demand for us to fulfil our sacred duty to consume. Engineering might ship a dozen improvements to approvals in the time it takes marketing to put a campaign together, and even if every release note is a masterpiece, the product they sat down to explain on Monday has moved on by Friday. If they keep the campaign current, they spend half the week revising work they’ve already done, and even if they can keep up, there’s the small matter of how often a customer wants to hear about approvals. At some point marketing has to decide which changes are worth asking the market to learn, and how often.
That campaign gets sales a conversation. A prospect sees the feature, tries it and asks what happens if an editor changes a date. Great. We’ve learned something, and engineering can have another version ready tomorrow. But suppose the prospect says, “Yes, that does what we need. Now our IT director has to approve it.” The economics change halfway through the call. Yesterday, a day of engineering could move the deal forward. Today, the person who can move it forward is the IT director.
Then the customer buys it, which is where the cost of change gets interesting. Before launch we can talk about our feelings and experiment with substances as much as we want. Once a customer has trained two hundred publishers, written guidance and agreed with Legal what counts as an approved revision, our next bright idea has a price. Their administrator might understand the improvement in ten minutes. The expensive bit is changing what two hundred other people do. Real-world use can accelerate the builder’s learning at exactly the point when the customer starts to demand stability.
How to apply this
Take one feature to the next product review and follow it all the way to the person who’s meant to benefit. Where did it spend the last fortnight? Suppose the prototype took a day, the team waited nine days for an editor to try it, and the editor settled the question in twenty minutes. You had the engineering resource. The missing knowledge was cheap to create once the right two people were in a room. Having the engineer join an existing customer call could save more time on the next feature than any improvement to the coding process. That’s what an arrangement problem looks like in practice.
Of course, the customer might have tried the thing, liked it, and asked us to hold off until next quarter because that’s when they’ve already got two hundred publishers in a room for their annual policy training. On our dashboard it’s another feature sitting around waiting for rollout, but bringing it forward a month means someone on their side has to organise a second round of training so we can say we shipped it sooner. I’d wait too.
I’d give teams room to build as quickly as they can learn from what they build. I’d be much more deliberate about how often we ask everyone else to change. Digital editing gave a director the freedom to try five openings, watch them back and discover that the third one made the film. The audience got one opening. Making five was worthwhile because only one had to leave the editing room.