Testing our economic framework against real-world agent team results from Opus 4.6 deployments.
Links: Economics, Value and Profit, Risk and Entrepreneurship, LLM Grounding Problem, Claude Opus 4.6 Research, LLMs as Praxeological Actors, The Anchor Method — the practical playbook this theory underwrites
We built a theoretical framework covering:
This page tests that framework against the real-world success stories reported in the Opus 4.6 video.
What happened: 16 Opus 4.6 agents built a fully functional C compiler in Rust over 2 weeks. 100,000+ lines of code, builds Linux kernel on 3 architectures, passes 99% of torture test suite. Cost: ~$20,000.
Structure: Lead agent coordinated, specialists handled subsystems (parser, code generator, optimizer), peer-to-peer messaging between agents.
Utility/Trade: Compute cost ($20K) exchanged for a product (working compiler). A human team would cost hundreds of thousands and take months. Massive utility gap — both sides profit.
Risk: Low. A compiler has a specification — it either compiles correctly or it doesn’t. Success criteria are clear and measurable. No entrepreneurial bet on demand or market. This is COO-class work — well-defined, verifiable, structured.
Grounding: Largely a non-issue. Code is text. Compilation is verifiable. The agents work in their native medium and outputs are testable in the same domain.
Verdict: COO-class problem. Agent teams excel here. No CEO judgment needed, no grounding gap.
What happened: Opus 4.6 on their issue tracker in production. Closed 13 issues autonomously in one day. Assigned 12 issues to correct team members across 50-person org and 6 repos. Knew when to escalate to a human.
Structure: Single agent operating as COO — routing, triaging, assigning. Building toward 24 parallel agent sessions on their monorepo.
Utility/Trade: Replaces coordination labor. A human engineering manager doing ticket triage spends 15-20 hours/week on pattern-matching work. At $250K/year fully loaded, that’s significant salary for automatable tasks.
Risk: Moderate. Closing issues = judgment calls on correctness. Assigning issues = predicting who has context. But critically — it knew when to escalate. It didn’t try to play CEO. It handled COO-level decisions and punted the rest.
Grounding: Small gap. Code and issue trackers are text-native. Org chart understanding works when organizational structure is documented. Risk would emerge if the agent tried to infer undocumented social dynamics.
Verdict: Clean COO-class success. The critical design choice: human-in-the-loop based on stakes. The agent presents options for hard calls rather than making them unilaterally. Exactly our model.
What happened: Given basic tools (Python, debuggers, fuzzers) and pointed at open-source code with no specific instructions. Found 500+ previously unknown zero-day vulnerabilities in code reviewed by human researchers and scanned by automated tools. Independently decided to analyze git history when standard methods failed. Invented its own detection methodology.
Structure: Solo agent with tools. No team, no hierarchy.
Utility/Trade: Finding vulnerabilities human researchers missed. Compute cost vs. security for millions of users. Enormous utility gap.
Risk: Here’s where it gets interesting. The agent made a novel methodological bet — “standard fuzzing isn’t working, I’ll try analyzing git history instead.” That’s not constraint execution. That’s closer to entrepreneurial judgment — a creative gamble that paid off. But the domain has low downside: if the novel approach fails, you just don’t find anything. You don’t lose resources.
Grounding: Partially relevant. Code is text, but vulnerabilities exist at the boundary between what code says and what it physically does (buffer overflows, race conditions). The agent reasoning about “where assumptions break down” is reasoning about the gap between text and physical behavior. Going to git history — reasoning about the code’s evolution over time — goes deeper than static text analysis.
Verdict: Pushes beyond COO into limited CEO territory. Entrepreneurial in method, but low-risk in outcome. A notable demonstration of creativity — the new model showing capacity for novel problem-solving approaches when standard methods fail.
What happened: Two non-technical reporters built a project management dashboard (calendar views, email integration, task boards) in under an hour. $5-$15 compute cost.
Structure: Human-agent pair. Human describes outcome, agent executes.
Utility/Trade: Personal software worth thousands in subscription fees for $15 in compute. Massive utility gap. But this is personal software — no repeatable value creation, no customers, no business risk.
Risk: Near zero. Bad output = wasted $15. No market bet, no demand risk, no competition. Below COO-level — individual task execution.
Grounding: The human closes the loop. They know what they need from physical experience using Monday.com. They describe it in text. The agent builds it. The human validates whether the output matches real-world needs. Grounding provided externally.
Verdict: Below the business framework entirely. A single trade, not a business. But illustrates a key principle: when the human provides grounding and judgment, the agent just needs to execute. Bottleneck is clarity of intent, not agent capability.
| Example | Decision Class | Risk Level | Grounding Gap | CEO Needed? | Outcome |
|---|---|---|---|---|---|
| C Compiler | COO — structured, verifiable | Low | None (code is text) | No — specification exists | Success |
| Rakuten | COO — routing, triaging | Moderate | Small (documented org structure) | No — escalates when needed | Success |
| Security | CEO-adjacent — novel methodology | Low downside | Partial (code vs execution) | Limited — bounded risk bets | Success |
| CNBC Reporters | Below COO — task execution | Near zero | None (human grounds it) | No — human is the CEO | Success |
Every success story either:
None of them required genuine CEO-level entrepreneurial judgment under high-stakes uncertainty.
The closest — security research — was entrepreneurial in method (creative problem-solving) but low-risk in outcome (bounded downside). This is notable: LLMs may develop creativity before they develop entrepreneurial judgment, because creativity under low stakes doesn’t require the slow feedback loops that make the CEO role hard to train.
The open frontier — running an actual business with:
The vending machine was the closest attempt, and it failed exactly where our theory predicts: at the intersection of sycophancy, grounding, and entrepreneurial judgment.