SOFIA is a method for working with specialized AI personas, in intentional friction, steered by a human who arbitrates.
OPEN SOURCE · MITTable of contents
This book was born from an observation and a frustration. It's an opinionated practitioner's synthesis — referenced, unapologetic — not a white paper: a blue book.
The observation: generative AI is reshaping the field. Not slightly — fundamentally. Those who ignore it will waste time. Those who blindly trust it will waste more.
The frustration: the problem with LLMs is structural[1] — a system that predicts the next token doesn't have a model of the world, it has a probability distribution — and the dominant discourse refuses to face it. That discourse offers only two stances. Replace people, or dig in your heels. As if there were nothing between full automation and refusal.
There is something else. A third way, built in the field — not in a pitch deck. It rests on a simple intuition: friction — between human and machine, and between machines themselves — is not a problem to solve. It's the mechanism that produces value.
What I describe here is not a theory. It's a method tested — on a real project, with real constraints, by someone working alone with limited resources. The results are there. The limits too. Both are documented.
This document is subject to the method it describes. It was produced with friction, challenged by constrained roles, and its limits are documented here — not hidden. Criticism is welcome — that's the mechanism. The repo is open: github.com/oxynoe-dev/sofia
This method was built empirically — one project, one practitioner, 210+ sessions. It's not a controlled study or a protocol validated at scale. It's a documented practice, subject to error and approximation. What it claims, it can demonstrate on its own ground. Beyond that, everything remains to be proven.
If you're looking for a magic productivity promise, this is the wrong book. If you're looking for an honest method to go further without losing control, you're in the right place.
Fragment I
The problem
A standalone LLM says yes. Always.
It codes, advises, writes — in the same conversation, with the same tone, unconstrained. It challenges nothing. Ask it a poorly framed question, and it will produce a well-formulated answer. Give it a shaky direction, and it will execute with enthusiasm. That's not collaboration. That's servile execution.
And yet, that's exactly what the market pushes. More automation. Fewer humans in the loop. Agents that do the work, people who supervise. The pitch is simple, the dream is clean, the demos are impressive.
The problem lies in the arithmetic nobody wants to look at.
10 serial steps — 90% reliability per step
Cumulative reliability = 35%
An agent that's 90% reliable on a single step — that's fine. Ten steps in sequence, the overall error rate climbs to ~65% (1 − 0.9¹⁰ ≈ 0.65). The error from step 2 enters step 3 as a valid premise. Step 3 builds on it. The cascade is silent. The final output looks correct. It isn't. This calculation assumes independent errors — in practice, correlation between steps can make things worse.
Salesforce saw it in production: beyond a handful of directives, LLMs start dropping some — Agentforce's CTO cited an empirical threshold around eight[12]. Cemri et al. (2025) analyzed 1,642 execution traces across 7 multi-agent frameworks: 14 failure modes identified, split between design problems (41.8%), inter-agent misalignment (36.9%), and task verification (21.3%)[13]. Our intuition is that multi-agent without governance degrades reliability rather than improving it.
And the nature of the mechanism is structural[1]. A system that predicts the most probable next token cannot be made reliably factual — its errors compound exponentially with sequence length. This isn't a bug to fix in the next release — it's how the technology works. Building massive automation on that foundation means stacking uncertainty on uncertainty.
What the demos
don't show
Demos show magic prompts that produce code in 30 seconds. What they don't show: the years of context in the head of the person prompting. The prompt is just the surface. The depth is everything that came before.
AI amplifies. It doesn't invent.
Feed it nothing, and it produces well-formulated nothing. Feed it years of conviction about a real problem, and it builds with you. It's a mirror — it reflects what you bring. Good framing, clear direction, real question: enrichment. Fuzzy framing, weak direction, poorly asked question: convincing confusion. And convincing confusion is more dangerous than a clearly wrong result — because you don't see it.
That's the hidden condition of value. The profile that gets the most out of AI isn't the one who codes faster. It's the practitioner who already understands their domain — who knows which questions to ask, and who uses AI to hold complexity at a level of detail they couldn't reach alone. A software architect with ten years in the field. A doctor who knows their edge cases. A lawyer who knows where the text breaks down. Domain expertise is the prerequisite — regardless of duration, it's depth that matters.
I see it on my own ground: 18 years of thinking about one specific problem — AI doesn't start from zero with me. It starts from where I am.
It's not "AI does the work for me." It's "AI lets me work at a level I couldn't reach alone."
A qualitative difference, not a quantitative one.
Fragment II
The thesis
It is possible to go faster and better with the same headcount — not at constant total cost, the transfer of load to infrastructure is real (see §V).
Not fewer people. The same people, augmented. Not replaced — amplified. An architect who holds three levels of complexity in parallel because AI helps them not drop anything. A developer who explores four approaches in an hour instead of one. A strategist who tests their hypotheses against structured challengers before presenting them.
Growth with stable headcount. Shneiderman frames it: high automation and high human control can coexist — it's a matter of design, not a trade-off[22].
It's less sellable. It doesn't make for a spectacular demo. It doesn't promise to cut costs tenfold. But it's sustainable. Because the human stays in the loop. Because when things break, someone understands why. Because competence is maintained instead of eroding[5].
Friction is the value mechanism, not an obstacle to eliminate. Intentional friction and role isolation aren't methodological luxuries. They're the condition for value[23]. La Rosa and Beretta formalize this principle within the framework of joint cognitive systems: friction must be designed as a scalable design element, adapted to the functional role and degree of control of each actor in the system[28].
In the market, everyone is trying to reduce friction with AI. Fewer prompts, more autonomy, agents that work on their own. My approach is the opposite: I generate friction to move the product forward. Specialized AI personas. Each with a scope, constraints, a stance, and a duty to challenge the others. The architect says "not now." The researcher says "your reference doesn't hold up." The strategist says "nobody will pay for that." If all the personas agree, they're useless.
The human isn't removed from the loop — they're the only one who can resolve what the agents can't resolve among themselves.
The field already confirms the theory.
Klarna — 2024. The CEO announces that an AI chatbot replaces the work of 700 support agents. The press applauds. A year later, he admits quality collapsed — robotic responses, customers stuck in loops. Klarna restarts hiring humans[9].
Same pattern at IBM[10] and McDonald's[11]. A weak signal, not a proven pattern — but a signal that keeps repeating.
The pattern when it appears: AI can do the job → headcount reduction → edge cases pile up → people called back. According to Forrester (2026), an analyst firm, 55% of companies that laid off workers for AI-related reasons regret the decision[7]. A third of them reportedly rehired between 25% and 50% of the eliminated positions[8] — figures to take with caution, both sources are commercial.
Bainbridge predicted this cycle forty years ago[2]. The only difference with LLMs: speed. What used to take a decade with outsourcing now plays out in months.