CEO, VeUP · Naples, FL · AWS Partner
Client delivery runs on an agent swarm. The tooling is public and the numbers are audited.
Multi-Agent Delivery · Agentic Coding · Tropical Attention · AWS · Rust · Swift · Local-First AI
Operating Model · Open Source · Research · Not Public · Instrumentation · Principles
Blue is work I direct. Amber is work the swarm directs on its own — about half of it.
Note
83.2B tokens across 122 days, on 24 models from 9 labs, 93% of it served from cache. About 48% was spent by subagents rather than by me typing. These are not estimates — they are reconciled from the transcripts on disk by flightdeck, and this paragraph is regenerated hourly from the same data as the badges above.
I run an AWS partner business where delivery is increasingly executed by agent swarms rather than by hand. That is a claim people make loosely, so here is what it concretely means day to day.
| Scale | 123 sustained concurrent agent sessions at peak on 2026-07-28. Not a burst — sustained across a working day, against a shared work queue. |
| Fan-out | 48% of recent tokens are spent by subagents. Much of the work is delegated by other agents, not by me. |
| Provider diversity | 24 models across 9 labs. No single-vendor dependency, and routing is a decision the model makes, not a hardcoded table. |
| Cost discipline | 93% of tokens served from cache. Cache is the difference between a swarm being affordable and being a science project. |
| Accounting | Every token attributed to a model, an account, a session, and a cost — priced at the rate in force when it was spent, not today's rate. |
The interesting problems at this scale are not prompting problems. They are coordination, attribution, recovery, and measurement problems — which is why most of what I build is infrastructure rather than applications.
| Project | Lang | What it does |
|---|---|---|
| flightdeck |  |
Truthful multi-provider token accounting for local AI coding agents. Reads the transcripts Claude Code, Codex, Grok and anything behind a Claude Code shell already write to disk. Recovers months the tooling deleted, treats subagent spend as a first-class dimension, and prices every token at the rate in force when it was spent. |
| zfc |  |
Zero Framework Cognition (Yegge) as an agent skill. Keep judgment in the model and heuristics out of the application. Violation catalog, audit playbook, model routing, identity discipline. Every regex you write against model output is a bet against the next model. |
| claude-cli-oauth-brain-transport |  |
Using the Claude CLI as a first-class inference path for swarm orchestration. What the print-mode envelope actually guarantees, why --json-schema does not constrain it, and how to get reliably parseable output anyway. |
Tropical attention at the edge. Replacing softmax with the max-plus (tropical) semiring, where a ⊕ b = max(a,b) and a ⊗ b = a + b. Every multiplication collapses into add-and-compare, which changes what the hardware has to be:
- Structurally predictable latency. Re-bracketing attention as
(V ⊗ Kᵀ) ⊗ Qnever materializes the L×L matrix, so it maps onto an add/compare systolic array. Sub-millisecond determinism becomes a property of the algebra rather than something you measure and hope for. - Full fidelity. No floating-point accumulation, so results are exact rather than approximately reproducible.
- Interpretability that is not a heatmap. The argmax at each node is a discrete route. You get genuine attributions and an ℓ∞ robustness certificate of radius ≥ min-gap/2 — a proof, not a saliency picture.
Built on Jeffrey Emanuel's model_guided_research, which explores eleven exotic mathematical structures as drop-in attention mechanisms. Carried into on-device safety inference and a scoring engine.
Most of the work is not open source — it either encodes something client-specific or is not finished enough to hand someone. The shape of it:
| Swarm orchestration | Running agents by the hundred against a shared work queue: leases, federation, recovery, and the accounting that proves what they actually cost. |
| Meeting intelligence | Local-first Mac capture. Transcription and diarization on-device; audio never leaves the machine, only text reaches a model. |
| A unified tool server | One MCP surface over the dozen-odd business systems delivery actually runs on, so an agent gets a single contract instead of a dozen auth dances. |
| Autonomous operating reports | Agents that assemble the monthly estate review, where every claim has to link to a receipt or it does not ship. |
| Delivery practice | Forward-deployed engineering, AI-DLC, and living roadmaps that stay current because agents maintain them rather than people remembering to. |
What generalizes gets extracted and published here. What encodes a client stays in.
The badges above are not hand-typed, and the story of why they exist is the reason I trust them.
Every usage dashboard on my machine was wrong in a different way. One was reading a cache that had not updated in four months. Another counted characters instead of tokens. All of them either double-counted subagent transcripts or could not see them at all. So the reported estate was ~35B tokens.
The audit put it at 81.7B — the figure above is higher because it keeps moving. Three things had gone wrong:
- Claude Code deletes transcripts after 30 days by default (
cleanupPeriodDays). April and May were simply gone from disk. The tell was that a cumulative counter had fallen — 56.62B in May, 46.90B in June. A total that decreases is proof of deletion, not of lower usage. - "Per-event data always wins" discarded 11.78B of real subagent burn, because archive sources sometimes hold sessions the per-event stream never captured. The correct rule is max-per-session: both sources are floors.
- 22% of tokens were priced at $0 because they matched no pricing pattern, and one model was seeded 4× low.
flightdeck exists so those failures are detectable rather than silent. It reconciles across sources, prices with effective dating, and reports the percentage of tokens matching no pricing rule so a silent $0 can never masquerade as thrift.
Structure in code, judgment in the model. The application is dumb pipes; the model is the smart endpoint. Schema validation, budget caps, and retry policy belong in code. Ranking, classification, and "is this done" belong to the model. Every heuristic is a bet against the next model release, and that bet has lost every quarter so far.
A number you cannot audit is a number you cannot manage. Most of this industry is reporting AI usage figures nobody has checked. Mine are checkable, which is the entire reason the collector is public.
Receipts or it did not happen. Internal reporting rule: every claim links to a diff, a test, or a demo. Prose that cannot cite itself gets cut.
Identity comes from email, never display name. Fuzzy matching is a last resort, never an entry point. An unresolvable entity stays unresolved rather than minting a junk canonical ID — the alternative is a corpus that quietly grows 290 records into 1,214.
I run VeUP, an AWS partner. The work I care about is at the frontier: delivery executed by agent swarms, measured honestly enough to trust.
