How One Person Can Run a Medical Software Company

Introduction

Something changed in software over the last eighteen months, and most of the industry is still catching up to it. AI models stopped being autocomplete and started being colleagues. They went from suggesting the next line of code to completing multi-hour engineering tasks, remembering what they did yesterday, and coordinating with each other.

Medware saw it coming and rebuilt itself around it. The result is a working system, in production today, in which a regulated Australian medical software company runs more than twenty products with three people, and the entire software development function is carried by one of them. The development team is not outsourced and it has not disappeared. It has been replaced by an organisation of AI agents, built on Anthropic's Claude, coordinated through three applications: Nebula, Constellation and Mission Control.

This article explains the problem that forced the system into existence, the opportunity that made it possible, how it works, and, as honestly as we can, what it is worth.

Background: the problem

Medware's founder has spent over three decades across pharmaceuticals and medical software. His previous company, Interact, ran with a team of more than thirty. Medware itself used to operate with seven people plus consultants. Today it is three people plus advisors: the founder, who covers development, business and sales; one person across finance and marketing; and one person managing PBS and clinic liaison.

That contraction is not unusual. It is the standing reality for most Australian health-tech companies, which live inside a brutal equation. On one side, the work never shrinks: TGA obligations, PBS reimbursement rules, clinical advisory input, security, privacy and data sovereignty expectations around patient information, plus all the ordinary demands of building, documenting, testing and supporting software. A traditional small software team capable of covering that ground costs serious money. Australian developers average roughly $85,000 to $130,000 a year depending on seniority, project managers around $140,000 to $160,000, with QA, design, technical writing and support roles behind them, and the true cost of each employee lands at 1.25 to 1.4 times base salary once superannuation, leave and overheads are counted. A modest six-to-eight person product team is realistically a one-million-dollar-a-year commitment before you have shipped anything.

On the other side of the equation sit revenues that, in specialised medical niches, often cannot support that headcount across twenty products. Something has to give, and historically what gave was either scope (build less), quality (document less, test less) or the founder's health (do everything yourself, badly, at 2am).

There is a further dimension in this case. The founder has ADHD. Running twenty-plus products solo means holding twenty schedules, four email accounts, dozens of repositories and a constant stream of regulatory tasks in working memory, and working memory is precisely what ADHD compromises. The standard advice scales poorly: research at UC Irvine by Gloria Mark found it takes an average of about 23 minutes to fully refocus after an interruption, and the American Psychological Association summarises task-switching research as costing up to 40 per cent of productive time. For a brain that switches contexts involuntarily, an unstructured day is a tax on everything.

Background: the opportunity

What changed is the capability curve of the models themselves, and it changed fast.

In May 2025, Anthropic's Claude Sonnet 4 solved 72.7 per cent of SWE-bench Verified, the standard benchmark of real-world software engineering tasks drawn from actual GitHub issues. By November 2025, Opus 4.5 became the first model past 80 per cent. By May 2026, Opus 4.8 reached 88.6 per cent, and in June 2026 Anthropic released Claude Fable 5, its first Mythos-class model, which roughly doubled the previous state of the art on the hardest frontier coding evaluations. In twelve months the benchmark moved about sixteen points, on top of a near-doubling the year before.

Endurance moved just as quickly. In mid 2025 the headline was Claude working autonomously for seven hours on a production refactor. By late 2025 Sonnet 4.5 was sustaining coherent work for more than thirty hours. METR, the research group that measures how long a task an AI can complete independently, finds that this task horizon has been doubling roughly every four to seven months.

Meanwhile the economics inverted. Anthropic cut its flagship pricing by about two thirds in late 2025, prompt caching now discounts repeated context by around 90 per cent, and small models such as Haiku 4.5 deliver near-frontier coding quality at a third of the price of mid-tier models and a tenth of the flagship. Anthropic's own data puts typical Claude Code usage at around US$100 to US$200 per developer per month. Compare that with the fully loaded cost of the human team above and the opportunity is not subtle: the marginal cost of competent software labour has collapsed by two to three orders of magnitude, for whoever can organise it.

That last clause is the catch, and it is where Medware's system lives.

What Claude still cannot do alone

Be clear-eyed about the gaps, because the architecture only makes sense as a response to them.

First, models are stateless. A Claude session does not natively remember the previous one. Left unmanaged, agents forget decisions, confuse one project with another, and write code into the wrong repository. Anthropic has shipped genuine progress here, including a memory tool and automatic context compaction, but its own documentation recommends external memory precisely because summarisation is lossy.

Second, context degrades. Research by Chroma across eighteen frontier models found that every one of them loses accuracy as input grows, with degradation observable from around 50,000 tokens even though windows now reach a million. A bigger window is not a bigger brain; it still pays to feed models small, correct, current context rather than everything you have.

Third, reliability decays with task length. METR's measurements are of 50 per cent success horizons, which means that at the frontier of what an agent can do, it fails about half the time. Independent research suggests agent success decays roughly exponentially as tasks lengthen. Long autonomous runs need checkpoints, review and verification.

Fourth, models still confabulate. The best grounded-summarisation hallucination rates are now under five per cent, but open-ended tasks remain worse, and in a medical software company a confidently wrong compliance claim is not a quirk, it is a liability.

Fifth, regulated domains demand human oversight by design. The EU AI Act's human oversight provisions for high-risk systems take full effect in August 2026, and Australian expectations in health are heading the same way. An AI software company in medtech must be built so a human can see, steer and stop everything.

Every component of Medware's system maps to one of these five gaps. That is the design insight: do not wait for the models to be perfect; build the scaffolding that makes today's models dependable.

The system: three applications and an engine

Nebula: the daily cockpit

Nebula is the founder's day-to-day app, on phone and desktop. It connects to all four company email accounts and calendars, takes notes by voice or text, and continuously answers one question: what should I be doing right now?

It offers three views of the same truth. A day view lists exactly what to do today; items are ticked off, corrected or expanded, and a spoken sentence becomes a properly filed to-do with reminders attached. A table view gives the overview. And a mind map lays the whole business out as connected nodes: each company, the products beneath it, the projects beneath those, with anything actionable hanging off its node as a note that updates itself and surfaces reminders.

Constellation: the single source of truth

Constellation is the registry of every project Medware runs: which local folder, which GitHub repository, which infrastructure, the schedule, and a running ledger of facts about what has been done and what is in flight. It exists because of gap number one above. Before any agent touches a project it reads the Constellation entry; when it finishes, it writes its updates back. Claude never goes to the wrong repo and always knows what it is up to, because the memory lives in the registry, not in any single conversation. The founder can add facts by hand, but Constellation is almost entirely maintained by Claude itself.

Mission Control: the software company in a window

Mission Control is the workforce: a full software organisation rendered as AI agents, each with a name, a role, a model behind it and a set of skills. The founder speaks to one agent, Atlas, who runs on Fable, the most powerful model Claude offers, and coordinates the heads of fourteen departments: project management, customer service, clinical advisory, security, legal, data analytics, QA testing, back end, front end, design, training, documentation, strategy and business operations. Senior agents such as Grace run on top-tier models; workers such as Oliver run on the smallest models and do the grunt work.

That tiering is a direct answer to the cost gap. Frontier intelligence is spent on judgement, coordination and review; volume work flows to models costing a twentieth as much. Work enters through an inbox: drop in a TGA compliance audit pack and the right departments engage automatically, pulling compliance and regulatory material from a shared library so nothing is lost or stale. Drop in a bulk PBS reimbursement feature and the system shows exactly who is involved, from design through clinical advisory, before the work begins. The whole team can be replicated per project, five, six, eleven times over, and viewed as an org chart or a network.

Claude: the engine

Underneath all three applications is Claude, and the discipline that makes it work is the read-in, write-out loop. Every agent, on every task, reads the right context from Constellation on the way in and writes its updates on the way out. Nebula talks to the emails, the communications and Constellation; Mission Control's teams talk to Constellation too. The three applications form a closed loop that keeps itself current. The founder does none of it by hand.

Why Nebula alone is worth the build

It would be easy to treat Nebula as the least impressive part, a to-do app next to an AI workforce. The evidence says otherwise.

Entrepreneurs are dramatically more likely to have ADHD than the general population. In Michael Freeman's well-known UCSF study, 29 per cent of entrepreneurs reported a lifetime history of ADHD against 5 per cent of a matched comparison group, and large-scale work by Wiklund, Lerner, Verheul and colleagues confirms that people with ADHD are more likely both to intend to start and to actually start businesses. The traits travel together: the restlessness, risk appetite and hyperfocus that build companies arrive with the working memory deficits that lose invoices.

The cost of leaving that unmanaged is not small. Deloitte Access Economics put the social and economic cost of ADHD in Australia at $20.42 billion in 2018-19, roughly $25,000 per affected person per year, with productivity losses exceeding $10 billion as the largest component. Medication helps but does not stick for everyone; population studies find only around half of adults remain on ADHD medication a year after starting.

What does the clinical literature recommend instead, or alongside? Externalisation. Russell Barkley, the field's leading researcher, argues that ADHD is not a disorder of knowing what to do but of doing what you know, and that the core intervention is converting working memory into physical, visible representations placed at the point of performance: the list where the work happens, the reminder at the moment it is needed, the map that shows the whole picture at a glance.

Nebula is that prescription, implemented in software and maintained by an AI. The day view is the externalised list. The reminders fire at the point of performance. The mind map is the visual field that replaces the mental juggling act. And because Claude maintains it automatically, it sidesteps the classic failure mode of every ADHD system, which is that maintaining the system is itself an executive function task. The systems that fail are the ones that need discipline to upkeep. Nebula needs none, because upkeep is part of every task the AI performs.

For one founder, the arithmetic is direct. A conservative reading of the interruption research, say one avoided derailment per working day at 23 minutes of refocus time, recovers roughly two working weeks a year. The realistic number for an ADHD founder running twenty products is far higher, before counting the missed deadlines, late fees and dropped follow-ups that the ADHD literature wearily calls the ADHD tax. If Medware had built nothing but Nebula, the project would already have paid for itself.

Quantifying the value

Now the whole system. The honest way to do this is to show both sides of the evidence.

The optimistic case: the original GitHub Copilot randomised trial found developers completed a well-specified task 55.8 per cent faster with AI assistance, and Anthropic's Economic Index shows coding dominating Claude usage, with around three quarters of enterprise API traffic following automation patterns, the model doing the task rather than advising on it. The sceptical case: a METR randomised trial in mid 2025 found experienced open-source maintainers were actually 19 per cent slower with the AI tools of that moment on large, mature codebases they knew intimately, even though they believed they had been faster. The lesson Medware drew is that AI does not speed up a human doing their old job; the gains come from reorganising the work so the AI does the job, with the human directing. That is exactly the difference between using a coding assistant and running Mission Control.

So compare the structures. The function Mission Control performs at Medware, development, QA, documentation, design, support triage, project coordination, maps onto a six-to-eight person team that in Australia costs around $0.9 million to $1.2 million a year fully loaded (a derived estimate from published salary bands and standard on-costs, not a quoted figure). The AI organisation performing that function runs on Claude at a cost measured in hundreds of dollars per month at typical usage, and in the low thousands in heavy months. Even being deliberately ungenerous, multiplying observed heavy-user spend several-fold for multi-agent overhead, the annual cost sits around one to two per cent of the human equivalent. The remaining 98 per cent is not all profit; it buys the founder's time as director, reviewer and final judge. But it moves software economics from headcount to tokens.

The market is reaching the same conclusion. Sam Altman has predicted a one-person billion-dollar company. Base44, a solo-owned startup, sold to Wix for US$80 million within six months of founding. Pieter Levels runs a multi-million-dollar product portfolio with no employees. What Medware adds to those stories is the demonstration that the same shape works inside a regulated industry, where clinical advisory, security, legal and audit-ready documentation are not optional, provided the system is built with the agents' weaknesses, not just their strengths, in mind.

Conclusion

Medware's AI development solution is three applications and a discipline. Nebula keeps the human pointed at the next right thing, and is on its own a clinically grounded answer to running a company with ADHD. Constellation gives a stateless technology a permanent, accurate memory. Mission Control turns raw model capability into an organisation, with hierarchy, departments, cost tiers and an inbox. And Claude, improving on a curve that has redrawn the benchmarks every few months, supplies the intelligence underneath.

The problem was a real one: a regulated software portfolio too heavy for a small company's payroll. The opportunity was real too: competent software labour suddenly priced in tokens. The solution is what sits between them, and it is the part that does not come off the shelf. The models everyone has. The organisation, the memory and the cockpit are the moat.

Three people. Twenty products. An entire software company, with one person where the dev team used to be.

---

Sources and further reading

Model progress: Anthropic announcements for Claude 4 (May 2025), Sonnet 4.5, Opus 4.5 through 4.8 and Claude Fable 5 (anthropic.com/news); SWE-bench Verified progression 72.7% to 88.6% via Anthropic and third-party benchmark trackers (Vellum, llm-stats). Task horizons: METR, "Measuring AI Ability to Complete Long Tasks" (metr.org). Context degradation: Chroma, "Context Rot" (research.trychroma.com). Coding productivity: Peng et al. 2023 (arXiv:2302.06590); METR developer RCT 2025 (arXiv:2507.09089). ADHD and entrepreneurship: Freeman et al. 2015; Lerner, Verheul and Thurik 2019 (Small Business Economics); Wiklund et al. meta-analysis 2026 (ETP). ADHD costs: Deloitte Access Economics for AADPA, 2019. Externalisation: Barkley, "ADHD, Executive Function and Self-Regulation" (russellbarkley.org). Interruption costs: Mark et al., UC Irvine (CHI 2008); APA, "Multitasking: Switching costs". Salaries: SEEK, PayScale and Australian IT salary guides, 2025-26. Claude pricing: platform.claude.com/docs pricing pages, June 2026. Solo-founder examples: TechCrunch and Wix press releases on Base44; Indie Hackers on Pieter Levels.