The Vorflux Manifesto, Read From Kolu
A study of vorflux.com/manifesto — the Great Flattening thesis, its six bottlenecks, tokenmaxxing — mapped honestly onto what kolu already is, what it should adopt, and where its doctrine disagrees.
A reading of the Vorflux manifesto (“The Great Flattening”), kept honest in both directions: where kolu is already the thing it describes, and where kolu’s doctrine deliberately disagrees.
What it claims
- The models woke up. Frontier models are “genuinely superhuman” at programming (SWE-bench Verified 33%→88.2% in two years; Codeforces 99.8th percentile) — and they’ve outgrown laptops, but “the cloud they were offered is blind”: agents can’t run and watch the application, so humans get dragged back for verification.
- The Great Flattening. Organizations collapse from layered hierarchy toward a harness — one system encoding company judgment. The human’s job “goes meta”: “you stop solving the task in front of you and start solving why the organism couldn’t solve it itself.” The human cell boundary survives at the outward-facing edge (sales, relationships); everything internal — planning, design, review, execution — flattens.
- Six bottlenecks on the way: the Machine (real multi-repo environments, not sandboxes) · Planning (a plan of plans) · Orchestration (route work to the right model per task) · Testing (live app testing, browser automation, recordings) · Review (cross-lab adversarial review — competitor models judging) · the Merge (automated conflict resolution + flags).
- Tokenmaxxing. “The seat is the wrong unit of compute. The token is the right one.” And once execution is cheap, backlogs are obsolete: don’t prioritize — “build them all and find out which ones matter.” Stay vendor-neutral across model families to avoid lock-in.
Kolu is already most of this — literally, this week
The manifesto describes, as a future, the shape of an ordinary kolu campaign. The six bottlenecks against kolu’s shipped stack:
| Vorflux bottleneck | kolu today |
|---|---|
| the Machine | padi hosts + the pu box pool — real repos, real daemons, leased per PR, egress-checked |
| Planning | the Atlas plan-of-record: id-first phase trees, dep-sequenced rows, gates, done-criteria |
| Orchestration | the coordinator’s model rule (“fable judges, opus grounds”), lens-run workflows, per-lane briefs |
| Testing | e2e + seal on the exact shipping SHA, two platforms; agents drive chrome-devtools against the live app |
| Review | the gauntlet — lens-debate, /agent-debate (an explicitly selected Claude, Codex, or Grok peer adversarially reviewing, warm-session, to consensus), code-police |
| the Merge | deliberately human — srid merges; see the disagreement below |
And the “job goes meta” line is kolu’s operating culture already: every coordinator failure this campaign became an encoded rule the same day (the atlas-branch liveness check, the /goal pin, the blocking-ask rule, the design-bearing lens trigger).All in /bridge and the surface rule, each carrying its recorded failure as the rationale — the manifesto’s “solve why the organism couldn’t solve it itself,” practiced as skill-writing. The manifesto’s sharpest observation — the blind cloud — names kolu’s exact differentiator: kolu is the anti-blind-cloud, a canvas where agents and the human watch the same live terminals, the same running app, the same dashboards.
What to adopt
- Say the thesis out loud. Kolu’s story is scattered across notes; the manifesto compresses it: the harness is the product. A kolu built for one user today is a company-judgment harness at any scale — the orchestrator skill, the gauntlet, the Atlas graph are the encoded judgment. Positioning, docs, and the website should claim this frame.
- Cross-lab review as a first-class feature, not a skill. /agent-debate proves the pattern with a chosen Claude, Codex, or Grok peer arguing to consensus. Generalize it further into a reviewer-panel primitive where N different-vendor agents debate a diff on the canvas — the manifesto’s bottleneck 5 as a product surface.
- Vendor neutrality as architecture. Kolu already runs mixed-vendor agents in PTYs — the neutral-harness claim is true by construction. Keep it structural: nothing in padi/kaval may assume one agent CLI (the
prohibitedKeybindsregistry is the pattern — per-tool facts as data, not baked-in assumptions). - Token accounting where seat accounting was. “Tokenmaxxing” needs meters: per-lane token/cost visibility on the dashboard and in reports (this week’s 81%-of-weekly-limit moment was discovered incidentally, in a terminal footer). A harness that routes work by cost needs cost as a first-class fact.
- Recordings as evidence. Bottleneck 4’s “recordings” names a gap: kolu’s evidence class is screenshots + test output; a padi-native “record this lane’s screen for the PR” would make visual evidence one verb.
Where kolu disagrees
- “Backlogs are obsolete — build them all.” Kolu’s doctrine says the opposite, with receipts: prove-then-extract, parked-with-gate, declined-dated-with-revive-triggers. SR10 was declined this week precisely because building it was cheap but owning it wasn’t — the constraint that survives cheap execution is not “what to build” but what to maintain, verify, and keep coherent. “Build them all” maximizes exactly the accidental complexity the design philosophy exists to kill. The roadmap-graph work keeps the gate discipline; it does not replace prioritization with spray.
- The fully-automated Merge. Kolu keeps the human at merge — not as a bottleneck but as the adjudication seat: this campaign’s record shows the human catching what every automated layer missed (the composition smell in a merged PR, a mis-clicked adjudication, a lying rules doc). The manifesto’s own “human cell boundary” is drawn at sales; kolu draws it at judgment — the flattening stops where accountability starts.
- Benchmark triumphalism. The 88.2%-therefore-superhuman framing is marketing shorthand; kolu’s culture is evidence-per-claim (a fix isn’t done at “CI green” but at real-GPU soak, live acceptance, exact-SHA proof). Adopt the thesis, not the epistemics.