The backlog
On 28 June I finished my last day as an employee. Three months later, 1,172 GitHub issues exist that didn’t exist then. I’ve closed The agents and I have closed 514 of them. I’ve taken the credit.
For most of my career that ratio would have been a warning sign. I now think it’s the most honest picture I have of what building with AI actually looks like.
Three months, in numbers
A little over 900 merged pull requests across about sixty repositories. Roughly 3,200 commits. Nineteen new repositories, from a shared platform that now runs on every Tessellant client’s server down to a weekend toy. Some of that is dependency bumps and my own build bots, so read it as an order of magnitude, not a scoreboard. (The first month on its own is in You can just do things, along with what it cost.)
Behind the numbers are things that actually run. In the first fortnight, three clients moved onto one common deployment. By the end of July every client repository had a test gate, and every deploy went through the same pipeline and rolled itself back on failure. A leagues club’s point-of-sale data now lands every night without anyone touching it. Last week a civil engineering firm asked whether AI could help check their drawings; three days later it was a working tool inside the chat system their engineers already use. In early September I built and deployed a complete platform for a new client in a single session. Most nights a build factory works through the queue on its own, and I read what it did over breakfast.
In week one I said a week was producing a quarter’s output. That held. It just turned out to be the less interesting number.
It’s worth saying who’s doing the building. My twenty years in IT started with a school-based apprenticeship and went through rapid-response support, consulting, support engineer, team leader, project manager, systems engineer and technical account manager, then service delivery manager, cyber security advisor, technical delivery manager and, most recently, IT manager. Most of it was in managed services. All of it was running systems, and looking after the teams and clients who depended on them, rather than writing software. I wouldn’t call this a career pivot; the systems background is the reason any of it holds together. But I’d never have started a business on my own, let alone as the one doing the building, without agents. Watching what they could actually do over the last year is where the certainty to start came from.
And “builder” undersells the job. I’m also the salesperson, the marketing department, the bookkeeper, operations and first-line support. In hindsight most of that is just what running a small business on your own looks like. The last month alone had a quarterly BAS, two NDAs, proposals and capability statements, blog posts and a platform upgrade to every client, alongside the building. It’s a lot.
The difference is that agents help with nearly all of it. Work that would once have eaten whole evenings - coding bills against the bank feed, preparing a contract for my lawyer, drafting the email that goes with a proposal - now takes minutes of my attention rather than hours. What the agents don’t replace is my lawyer and my accountant. If anything, they’ve made them matter more.
AI will get you most of the way on almost anything, and it sounds completely sure of itself the whole way. That’s the trap. It can still produce slop, and the slop reads exactly as confidently as the good work. My master services agreement was put together with a lot of help from agents. It read well and looked finished. My lawyer found a gap in it around work products - the kind of thing you only know to look for after years of watching contracts go wrong. Nothing in the draft hinted anything was missing, because the model had no way of knowing it was.
The agents make an adviser’s time go further, because what lands on their desk is clearer and better organised. They don’t replace the decades that let an adviser spot what’s missing. If you’re starting something on your own with AI doing a lot of the heavy lifting, a good lawyer and a good accountant are some of the best money you’ll spend.
Ideas were always cheap
What I hadn’t appreciated until this year is how much cheaper ideas get once something captures them properly.
July alone produced 666 issues - a number I’m choosing not to read into, although anyone who saw me in July might. One afternoon of staff interviews at a single client became about forty-five feature ideas, each written up with the problem, who raised it, and a rough shape. Every client conversation, every meeting transcript, every late-night “what if” turns into a well-formed issue in seconds. The capture is so good it’s a hazard.
By mid-July the backlog was inflating faster than anything could drain it, so I changed the rule. Every new idea is now born parked - labelled as future work, in a lane called Horizon - and only a deliberate review promotes it. The active lane is capped at nine items across every client. The count of open issues stopped being the metric: nine in flight, everything else deliberately waiting.
That brought capture down from 666 a month to around 200. It’s still more than I close. And the review that’s meant to do the promoting happens less often than the rule says - the active lane is over its cap as I write this.
Building got cheap too
The ideas pile used to be self-limiting, because building was expensive. You couldn’t do most of it, so you never had to agonise over most of it. The cost of building did the choosing for you.
That filter is mostly gone. An issue that would have been a fortnight’s work is often an afternoon. The drawing tool went from a question to something in engineers’ hands in three days. If an idea is well specified, the odds are good I can have a version running before the client has finished deciding whether they want it.
In week one I guessed that when building drops towards zero, the scarce thing becomes judgement about which output matters. That was a hunch after four days, and three months of backlog has borne it out. It sits a step before what I wrote in June, that verification is the scarce thing - you only get to verify what you decided to build.
Nearly everything in the backlog is something I could build. I just haven’t chosen to.
What’s actually scarce
Of the 492 ideas I’ve labelled as future work since June, 102 are closed. About one in five. The other 390 are sitting there, most of them perfectly buildable.
The pattern that stings is closest to home. Client work is urgent, visible and paid this month, so it wins the choosing almost every time. The longer-term work - the reusable pieces that should outlast any one engagement - only gets its turn when I stop and deliberately overrule the queue. The backlog showed me how rarely I’d been doing that.
When building stops being the constraint, the constraint moves. For me it’s gone to three places.
Judgement about which. Every idea in that pile has a plausible case. Most of them would work. In practice judgement looks less like picking winners and more like saying no to 380 good ideas so ten of them get finished properly - and saying it again next week, because the 380 don’t go away.
What the client can absorb. A forty-person firm can take on maybe one new way of working at a time. Ship them five tools in a month and you don’t get five times the value - you usually get less, because none of them sticks. A licence is not a deployment, and a shipped tool isn’t adoption either. The client’s capacity to change is the real rate limit, and no amount of build speed moves it.
Running what already exists. Nineteen new repositories is nineteen things that can break at 2am. Every tool that ships becomes something to monitor, patch, secure and hand over. I wrote in The ninety per cent that the build is maybe ten per cent of operating an agent inside a business. Three months in, that feels generous. The operations side has taught me enough to need its own post, so it’s getting one.
The same list, different weeks
Some weeks I’m sure it’s working. Some weeks I’m pretty sure I’ve ruined my life.
Nothing about the business changes between those weeks. Same clients, same pipeline, same bank balance, same backlog. What changes is how I read it.
On a good week, 682 open issues is a list of options - proof there’s more worth doing than one person can get to. On a bad week it’s a ledger of everything I haven’t done, every promise I made to myself in July and didn’t keep, and a quiet case that I bet the family’s income on a pile of ideas. It’s the same list. I’ve checked.
Early on I treated every dip as a verdict and re-ran the whole decision from scratch before lunch. What I’ve learnt is that the doubt is weather. It arrives on a Monday with no new facts in it and leaves by Thursday having changed nothing. The numbers are data, and they say the same thing on the good weeks and the bad: a lot got built, most of it is in use, and the urgent work has been winning the choosing more often than it should.
That last one is worth acting on. The rest blows over.
So, did I do heaps?
Yes. I’ll own that. More shipped in three months than I’d have believed in June, and most of it is running for people who use it every day.
The part I most want anyone reading this to take away is that none of it needed a team, funding or a computer science degree. It needed a subscription, a second-hand desktop in the garage, and a year of getting comfortable working alongside agents. It genuinely feels like the future turned up early. You can have an idea after dinner and a working version of it before bed, and turn that into something a client pays for in weeks rather than years. If you’ve been sitting on something, finding out whether it works has never been cheaper. You can just do things was about the permission. This is about the price.
If you do start, you’ll hit the backlog sooner than you expect. The ideas will come faster than you can finish them, and the building will keep up better than you’d believe. What’s left is choosing - a far better problem than the one most of us had a few years ago, when the hard part was everything.
Building got cheap. Choosing didn’t.