I am writing this on a Tuesday in mid-March, twenty-six weeks and roughly two hundred and seventy commits into building DuranteOS. Studio has been live for sixty days. The Hexagonal migration is essentially done. The procurement-week dispatch I shipped Friday reframed the architecture argument I have been making for two months as a sovereignty risk priced at the federal-procurement tier.
Six months is the right shape for a retrospective. It is long enough that the load-bearing decisions have had time to compound, short enough that I remember the alternatives I considered and rejected, and exactly the moment at which the founding-customer hunt's clock has burned half its time. The series itself runs for another seven weeks; this essay is the mid-series check-in, not the closing argument.
This is the twentieth essay in the series I started writing six weeks ago when I committed to making the case for DOS — the personal AI operating system I left a steady income to build — out in public, in detail, with the receipts. Twenty essays in is the right moment to step back, audit honestly, and decide what the back half looks like.
The half-year compressed into one sentence
The thesis held. The technical architecture held. The go-to-market plan was wrong in three concrete ways, two of which I have corrected and one of which I am still working on. The single biggest surprise was how much the writing helped.
What endured
Three load-bearing decisions made in the first month that I have not had to revise.
One. The Algorithm as the primary discipline. The Ideal State Criteria decomposition — turning every multi-step request into a numbered list of falsifiable claims before any work begins — has not been adjusted in twenty-six weeks beyond minor tuning. It is the most important load-bearing piece of DOS and the one I would point to first if asked "what is the actual product."
Two. The four-copy distribution model backed by a mechanical sync invariant. I designed it grumbling about the overhead. Five months later I would not change a single line of the manifest schema. The sync-check tool has caught seven distinct multi-copy drifts that would have shipped to customers. Worth every minute of the build cost.
Three. Builder's Compass as the only acquisition channel. I considered paid ads in month two and again in month four and decided against both times. The trust capital from over two years of weekly writing has done what no paid acquisition could do at the founding rate. The single most underrated decision of the half-year.
| Decision | Made in | Still standing? |
|---|---|---|
| Algorithm + ISC discipline | Month 1 | Yes — minor tuning only |
| Four-copy + sync-check | Month 2 | Yes — schema unchanged |
| Memory layer as SQLite knowledge graph (sketched, MemPalace fork pending) | Month 1 | Yes — design intuition matched the upstream that ships in April |
| Studio gateway (one Bearer token, multi-provider) | Month 3 | Yes — added providers, no architectural changes |
| Hexagonal Architecture for Studio | Month 5 | Yes — boundaries holding mid-migration |
| Builder's Compass as the only acquisition channel | Month 4 | Yes — still the most efficient channel |
| One reference customer at founding rate | Month 5 | Yes — still hunting, July deadline |
What changed
Three things I was meaningfully wrong about.
What I believed in month 1 vs what I believe now
- "Ten paying operators in the first six months" — too many, too shallow, would dilute the depth
- "I can sell DOS at standard pricing from day one" — wrong; the price band needed a founding rate first
- "Direct outbound to my consulting network will produce the founding cohort" — wrong; the cohort is coming from Builder's Compass readers, not from prior consulting clients
- One paying operator by July with a deep relationship is correct; ten would be wrong
- A founding rate is correct; standard pricing is the right price for the next phase, not the founding phase
- Public writing is the channel; the trust capital from BC readers is dramatically higher quality than the trust capital from prior clients
The first two corrections were applied in months 2-3. The third was the most expensive — I spent two months in late 2025 doing outbound to consulting contacts and produced approximately zero qualified founding-customer conversations. The pivot to BC-driven inbound was the second quarter's best decision.
What surprised
Three things I genuinely did not see coming.
The three biggest surprises of the half-year
- The agent's ability to debate itself usefully. The Council pattern was a side experiment. It became the single most-used decision-making primitive in DOS — far more than I projected. Multi-specialist debate, with the synthesis preserving disagreement explicitly, is the architectural review I always wished I had access to as a solo founder. I now run it on every non-trivial decision, even though the formal orchestrator is still a sketch (the informal version has been running for fourteen weeks).
- The compounding rate of public writing. I expected Builder's Compass to grow linearly. It grew super-linearly: 400 subscribers at end of year one, around 3,800 today. The mechanism — readers forwarding individual issues to other senior engineers — was visible in the data but I had not modeled how powerful it would become. The list is now my single most valuable professional asset.
- The substrate-bifurcation pace. I expected the Western frontier to consolidate and the open-weights tier to stay months behind. I was wrong on both. The bifurcation that hardened in W16 — Western labs at $380B priced as enterprise distribution, Chinese labs shipping frontier-class open weights on non-US silicon — was not on my month-1 map. The architecture I committed to in month 3 turned out to be exactly what survives the bifurcation, which is partly luck and partly the discipline of building toward portability rather than speed.
The numbers, honestly
| Metric | At month 1 | At month 6 | Where I expected to be |
|---|---|---|---|
| DOS paying customers | 0 | 0 (1 in active negotiation) | 0–1 |
| DOS skills/workflows | ~10 | ~80 | ~40 |
| DOS packs | 2 | ~12 (3 production-grade) | ~6 |
| Studio gateway operational | inactive | active (Altyaa via gateway, ~60 days live) | inactive |
| Builder's Compass subscribers | ~3,200 | ~3,800 | ~3,500 |
| Founding-cohort conversations completed | 0 | 11 | ~8 |
| Founding-cohort second conversations | 0 | 4 | ~3 |
| Posts published in this series | 0 | 20 | 12 |
| Failure-corpus signals (informal) | 0 | ~30 estimated | unknown |
| Personal runway remaining (months) | 12 | 6 | 6 |
The clock is the most important number. Six months of runway remaining means a binding deadline of roughly July that the founding-customer hunt has to clear. The negotiation in progress has to convert; if it does not, the next four months get the same intensity from the next two prospects. The runway number is the one I have not learned how to lie to myself about, and I am not going to start.
What I would not change
Three decisions that have been questioned (by readers, advisors, and my own occasional doubt) and that I would still make.
One. Leaving the consulting income when I did. The depth of work made possible by full attention is not approximate to half-attention; it is categorically different. I would have a worse product and a worse list if I had tried to do this part-time.
Two. Running Altyaa as the revenue line alongside DOS. The optimal-focus advice would say "shut down the second product." The optimal-focus advice is wrong here because Altyaa funds DOS's runway and validates the shared infrastructure (Studio gateway). The two products together are more interesting than DOS alone would be. The Studio launch in early January would not have made sense without Altyaa as the first internal customer.
Three. Public writing on DOS during the build, not after. The conventional wisdom says "build in private until you have something to show; then announce loudly." I did the opposite. The trust capital from public iteration through twenty essays has done what a single launch announcement could not. Writing was the marketing, the architectural-thinking forcing function, and the trust-building all at once.
What I would change if starting over
Two specific moves I would make differently.
One. I would invest in direct sales / GTM earlier. The current rebalance toward founder-customer development should have started in month two, not month five. The opportunity cost of the delay is at least three months of cohort-building — and it shows up directly in the founding-customer-conversation count being four under what it could have been with an earlier start.
Two. I would write the eval suite sooner. Building it in month six instead of month two meant four months of agent-behavior changes that I cannot precisely audit. Not catastrophic — the operator (me) caught most regressions — but unsupportable at scale. If I had a second founding customer right now, the eval suite would already be saving us both time. I am building it this month per the W16 essay.
The single discipline that mattered most
If forced to attribute the half-year's progress to one habit, I would name recurring-signal heuristic. The rule that any failure or correction repeating four or more times in 30 days requires structural fix has driven nearly every meaningful architectural improvement in DOS. The discipline is mechanical: count the signals, cross the threshold, do the work. The leverage is enormous, even running informally.
What I am betting on for the back half
The thesis remains: the agent is the product, intelligence replaces interface, memory replaces lock-in. The decade-shape of the bet is unchanged. What is in flux for the next six months is execution, in three specific lanes:
A small thank-you, named explicitly
A non-trivial number of readers will have made it through some of the twenty essays. A smaller number will have made it through all twenty. To both groups: thank you.
The writing was for me first — it forced the architectural intuitions into legible shape and the strategic bets into named claims. But the writing only worked because there were readers who would scrutinize it. Several of the essays were materially improved by reader replies: the memory replaces lock-in post benefited from a Brazilian DBA who pushed back on my confidence-decay heuristic, the Council pattern post sharpened after a former Stripe engineer questioned my non-convergence framing, and the Hexagonal Architecture post was directly improved by a reader who has been doing this longer than I have correcting two of my Cockburn citations.
The series is the audit. The audit is the asset. You helped build it. Thank you.
Where the series began
The thesis that frames everything.
The decision behind the series
The financial and identity bet.
The current chapter
The hunt that defines the next four months.
What continues weekly
The newsletter the writing extends from.
The series continues for another seven weeks. The work continues past it. The clock continues. The first reference customer is four months away on the schedule and not yet on the contract; that gap closes between now and July, or it does not.
Either way, this blog will keep producing the receipts. The next seven essays cover the build — Sentinel shipping, MemPalace forking once the upstream lands in April, the routing-bakeoff results, the first customer conversation if it converts, the post-mortem if it does not. The honest version of either kind of report is the only version worth writing. I will write them.
Until then, thank you for reading.
Was this page helpful?