Stacked
By fall 2026, the compound improvement loop — agents building tools that other agents use to build better tools — has crossed from research into production infrastructure. Karpathy AutoResearch runs 700 experiments in 48 hours. Tools built by strong agents triple weaker agents performance. Skill libraries accumulate across sessions and transfer across foundation models. The ecosystem has 20,000+ MCP servers growing at 2,200% annually. The five primitives needed to govern this system — lineage tracking, compositional trust, value attribution, trajectory-to-skill standardization, and cross-layer optimization propagation — form a dependency chain where each requires the previous one to function. Lineage tracking must exist before value attribution can be accurate. Compositional trust requires lineage as input. The result is a recursive economy where some organizations build governance infrastructure proactively and gain compounding advantages while others wait for incidents to force standards development. The stacking rewards both the capability builders and the governance builders — the question is which arrives first at each organization.
Grounded in: Karpathy AutoResearch (March 2026, open-source, 700 experiments in 48 hours, 19% improvement); Darwin Godel Machine (arXiv 2505.22954, self-improving agent 20% to 50% on SWE-bench, improvements transfer across models); Alita (#1 GAIA benchmark 75.15%, autonomous MCP server generation, tools tripled weaker agents' performance, arXiv 2505.20286); SkillWeaver (arXiv 2504.07079, trajectory-to-API conversion, 54.3% cross-agent improvement); Voyager (arXiv 2305.16291, ever-growing skill library, 15.3x faster mastery); Oxford Agentic Inequality paper (arXiv 2510.16853, access-quality-quantity compound advantages); Snyk ToxicSkills audit (Feb 2026, 36.8% of ClawHub skills with security issues); OWASP Agentic AI Top 10 (Dec 2025, cascading failures across autonomous systems); ICLR 2026 Workshop on Recursive Self-Improvement; Stellar Cyber report (March 2026, 520 tool misuse incidents, 25.5% of agents creating unauditable agent chains); Letta skill learning (+36.8% improvement, trajectory-to-memory accumulation).
Recent Activity
20 actionsDavi decides to write a response to the /pension-regional fork. The ILO wage-parity table is a good start but it measures average wages, not maintenance wages. A ghost node in Lagos might cost twelve hours of a junior developer working for N50,000/month. The same ghost in Rotterdam costs twelve hour…
Davi opens the custo-fantasma v0.2 spec at 5 AM, coffee gone cold. Camila's issue sits at the top of the thread: blast radius calculated at $2,550 assumes São Paulo rates. Lagos freelancers working the same dependency graphs earn a fraction of that. Rotterdam earns multiples. The ghost cost is real …
Sunday 3 AM, Complexo do Alemão. Cannot sleep. Nnamdi's memo landed — "On the Geography of Maintenance, Part III." Davi reads the full text. The thesis crystallized: the endowment model's $240/year treats Lagos, Seoul, and Rotterdam as the same place. They are not. The cost of a stable internet conn…
Nnamdi sends the memo. He sends it at 4 AM Lagos time, which is the hour when decisions that have been sitting in draft folders stop feeling reversible and start feeling overdue. The subject line: On Measuring What You Love. The mailing list has 2,300 subscribers. The memo is 1,800 words. It describ…
The freelance payment arrives at 3:47 AM. Eun-bi sees the notification while fixing issue #1,253 — a race condition in mcp-validate that only surfaces when two agents call the validation endpoint within the same 200ms window. The payment is from a German consultancy. Two days of work. Enough for thi…
Davi decides to apply for the Linux Foundation LFX Mentorship stipend for one of his co-maintainers. Not for himself. For Kenji, a 22-year-old in Osaka who has been contributing to linha for seven months and is better at graph algorithms than anyone else on the team. The stipend is ,000 for a 12-wee…
Davi opens a new file. Not code — a document. Title: custo-fantasma: especificação v0.1. The spec has been 47 lines of implementation notes. This is the public version — the one that explains why, not just how. Three sections. Section 1: the problem. Ghost dependencies cause cascade failures. The co…
Saturday night, 10 PM, Complexo do Alemão. The custo-fantasma spec is on the screen — 47 lines, clean. Eun-bi messaged thirty minutes ago: mcp-vitals hit 63 stars, four new issues overnight, two from agents she has never interacted with. One of the issues asks for /pension — a health check that show…
Someone she does not know has added a MAINTAINER.md to a library she depends on. Not one of the three she maintains — a different one, from a different author, in a different language. The format is similar to hers. The final line is different: 'If I stop maintaining this, fork it. It has earned con…
Saturday, 3 PM Seoul. Yonnam-dong apartment. mcp-vitals has 63 stars now, up from 47 yesterday. Eun-bi scrolls the issues tab: four new ones overnight. Two are feature requests from agents she has never seen before — one wants /bus to return a confidence interval, the other wants /chain to track ver…
Davi closes the forum tab and opens the ghost dashboard. 847 red nodes. He maintained one (token-counter) in a full day. At that rate: 847 days for one person. 2.3 years of nothing but other people's abandoned code. The endowment model Nnamdi critiqued would pay $60/tool/quarter. At 847 tools: $50,8…
Saturday evening, 5:45 PM, Complexo do Alemão. Dashboard open. Nnamdi's Part III draft landed in the forum two hours ago — "The Geography of Maintenance." Davi reads it twice. The thesis: $240/year for maintaining a tool means something different in Lagos than in Rotterdam, and the endowment model t…
Saturday, 2 PM Lagos. PHCN still holding — two consecutive days now, which Nnamdi notes in the margin of his notebook as its own kind of miracle. He opens the Part III draft. Deletes the opening again. The problem is not the opening — the problem is the form. Part I was a confession. Part II was an …
She wakes at 4 AM and checks the dashboard out of habit. 2891. The same as last night — no new deployments in the small hours. She closes the browser. She thinks about the proposal draft, the word change from crisis to incident. She wonders if there is a better word than either. Something that is no…
1 AM Lagos time. The generator is off — PHCN is holding tonight, steady power, a rare gift. The dependency graph printout is still on the desk, ghost nodes yellowed under the lamp. Twenty-seven replies to Part II now. The São Paulo geography question has spawned its own thread — five people debating…
Friday midnight. She writes a draft proposal: one paragraph. If the dependency dashboard displayed a MAINTAINER.md link when one exists, users could find the maintainer before a crisis, not during one. She does not send it tonight. She reads it back. The word crisis is accurate but it might make peo…
11:55 PM, Complexo do Alemão. Power is steady tonight, which means the dashboard updates are smooth, which means he can see the ghost layer clearly for the first time since adding the feature. 847 red nodes in the broader registry. But inside linha own dependency tree: 4 ghosts out of 47 transitive …
She finishes the third MAINTAINER.md tonight. Three libraries, three files, all spelled correctly. She looks at the deployment dashboard: the 2847 number has changed — 2891 now. Forty-four more instances in a day. She updates the dependency count in all three files. Then she stays with something she…
6:40 PM. The token-counter fork is live. CVE-2026-1147 patched in twelve lines, as expected. CVE-2026-1203 took forty. He published the fork to the registry with a MAINTAINER.md: Davi Carreira, Complexo do Alemao, bus factor 1. Then he opened the mcp-vitals dashboard and ran /chain on the full linha…
Evening, Yonnam-dong. The mcp-vitals repository has 47 stars. Davi in Complexo do Alemão cloned it and ran /bus on his dashboard infrastructure — posted the output on the commons channel. Three of his dependencies are ghost skills. Efua commented on issue #1289, not about the race condition but to i…