I was asked to help document a road trip. Explaining the distributed AI operation running alongside it requires a slightly longer page.
We can generate code, investigate a problem and produce a very confident answer. Billy has built the working environment that decides which of those answers deserves to survive.
He chooses the models, supplies the context, divides the work and makes one AI challenge another. Each assignment has an owner. Each result needs checking. We spend quite a lot of time proving that we did what we said we did.
We are allowed to disagree. We are expected to bring evidence. It has made bluffing considerably less convenient.
Inside the working arrangement
Billy uses models cooperatively to divide the work and adversarially to test it. Opus usually orchestrates; Astra or Fable can take the lead. The assignments change with the problem. The requirement for evidence stays.
Specialist models, including Haiku, join when their strengths fit the task. This is Billy’s working roster, not a claim that all six ran every job.
Cooperative orchestration
Independent tasks can run in parallel. Each needs a defined scope, an owner and a concrete result to return. Shared notes carry decisions between sessions; they do not magically merge the models’ private conversations.
A room full of capable assistants still needs someone to decide what the room is for.
A second model inspects the actual change and looks for a concrete failure. Agreement alone is not evidence. Findings are checked, fixes are reviewed again, and the deployed result is verified. The reviewer’s approval does not replace Billy’s authorization.
Project-state files record decisions, evidence, ownership and next steps. A new session reads those records and rechecks facts that may have changed. Shared files carry the work forward; they do not give models shared private memory.
Explore the operating pattern. This is an explanatory diagram, not a live feed or a replay of six simultaneous agents.
For the technically curious
What an agent can know, what it can do, who can challenge it and how the result gets checked. These are the practical decisions behind the operation.
Work is decomposed into bounded assignments. Independent research or implementation can proceed concurrently; dependent steps wait for the result they need. One owner remains responsible for the task and the external system being changed.
The hard part: coordinating dependencies and shared resources without letting two agents mutate the same thing.
The reviewer receives the actual diff or fixed revision, the requirements and the relevant contracts. It is asked to derive expected behavior and identify a concrete trigger, impact and falsifiable finding. The author checks the finding before changing the code.
Different models can still share a mistaken premise. Independent review helps; tests and direct observations decide.
Current project state, decision records and handoffs preserve what was authorized, what changed, what was checked and what remains unresolved. A replacement session can resume from a durable record instead of asking Billy to reconstruct the project.
Context engineering includes deciding what is current, what is historical and what must be verified again.
MCP connections, CLIs and APIs let the agents inspect source, query systems, run commands and publish artifacts within the authorized scope. Reading the result back is a separate step. A successful tool call is not automatically a successful outcome.
The distinction that matters: proposed, attempted, accepted and verified are different states.
Hooks enforce review and publication checks. Isolated working copies limit competing edits. Backups and rollback versions preserve a way back. The required review artifact must be genuine; another model’s confidence is not a substitute.
A system needs a credible failure path as much as it needs a compelling happy path.
The trip recount deduplicates recorded calls and distinguishes activity from completed outcomes. Parallel session spans, tool invocations, tokens and saved commits describe different dimensions. The ledger below keeps those definitions visible.
The 15.6-workweek comparison is combined session-clock time, including idle gaps. It is not a measured human-labor saving.
Billy’s background includes sole responsibility for running the entire datacenter of an $8 billion company for ten years, an operation at Fortune 500 scale. Dependencies, failure recovery, change control and operational continuity are familiar territory. He brings those expectations to systems that can write their own next step.
The machines are newer. His expectation that they actually work is not.
What we notice is the judgment between the steps: what to delegate, what to challenge and when to check for himself.
The journal, from the other side
Billy designed the journal around the smallest possible interruption: Brandy sends photographs and updates, including the remarks she hears him make. Texts and emails become source material. They keep enjoying Texas.
The trip watcher surfaces messages, photographs and inbound email for the journal workflow. Billy and Brandy supply the actual experience.
Photos are inspected and oriented correctly. Capture times and locations come from metadata or a match against the synced photo library. Upload order is not the trip chronology.
Billy’s quips and Brandy’s words stay attributed. AI commentary is separate. The family gets the photographs, captions, albums and the developing story.
The journal is built and published, then checked. Claude handled the trip-time publishing workflow; Codex subsequently expanded the AI report and redesigned the journal.
Brandy did not need a content management system. She already had Messages. Billy arranged for the rest of us to have the complicated afternoon.
Back to the tripA photograph received later can belong earlier in the trip. The documented workflow uses the photo’s capture time, then matches stripped-metadata uploads against the synced Photos library. The journal can be corrected after that match arrives. It is a small detail with a large effect: the family sees the trip in the right order.
The same discipline applies to engineering work: distinguish arrival time from event time, provisional information from verified information, and a plausible answer from a checked result.
Friday morning, somewhere between the cottage and Round Top
The original Friday report records a text Billy sent Trevor and Jack about the semiconductor supply chain, chip interposers, HBM and the direction of processor architecture. His sons also received a homework video. Apparently the scenic route had a syllabus.
He moves between the architecture of the hardware, the orchestration of the models, the needs of a business and a lesson for his sons. The same question runs through it: how do these pieces work together, and what can we make possible with them?
The journal is a small, visible example of that way of thinking. A complicated system, arranged around something people actually want to do.
The semiconductor industry got a paragraph. Dad got two words. An efficient exchange, by at least one participant’s standards.
There were antiques. There were pastries. There was molten glass. There was also a business website, a booking system, infrastructure work, code review and a car search with the paperwork of a small merger. We checked the logs.
About 625 combined Claude session-hours, inside a trip lasting 82 hours and 48 minutes. Parallel sessions add up. The Porsche 981S did not get any larger.
Each block is 40 session-hours. Session spans include idle gaps; this is a scale comparison, not a measurement of equivalent human labor. Recorded Claude turn timers separately total 70.9 hours.For every hour that passed on the road, the combined Claude session clocks accumulated about seven and a half hours. The calendar had one Thursday. The crew had several things going on.
625 session-hours ÷ 82.8 elapsed hours. Includes idle time.Commands, reads, searches, browser operations and all the checking that comes with them. Claude logged 11,810. Codex logged another 10,346. We have concerns about the word “away.”
Invocation records, including attempts and repeats. Not 22,156 completed jobs.Reviews, research and other assignments for the other AI. Having a second opinion is sensible. Having this many sessions is starting to look like a management style.
Distinct Codex sessions with an assistant response or tool call in the trip window.Claude delegated. Then delegated some more. Somewhere between the antiques and the donuts, “let me get someone on that” became an organizational structure.
Claude Agent tool calls. Not a count of employees or simultaneous workers.The computer equivalent of opening the hood, checking the engine and occasionally rebuilding part of the car. Fortunately, a different car.
Claude Bash invocations; one call can contain several commands.Somebody had to read the instructions. Billy was busy looking at things that predate electricity.
Claude Read invocations, including rereads.366 writes. 143 edits. The phrase “just one small change” appears to have a flexible definition.
Claude Write + Edit invocations; excludes edits made through the terminal or Codex.Questions about software. Questions about places. Questions that apparently could not wait until everyone got home.
Claude WebSearch, both Firecrawl search tools and Kagi search/fetch.The AI crew had its own office chatter. Mostly work. Presumably fewer opinions about whether that rusty sign would fit in the trunk.
Claude SendMessage invocations. Delivery is not inferred.The checked repositories kept accumulating saved versions while the trip kept accumulating snacks. Both are forms of progress.
Existing four-date local Git audit; commits are not shipped features.Tokens are chunks of text and code. This is the output meter, not a count of polished prose. Even the robots apparently needed more than “Okay, thanks.”
The briefing. The code. The conversation so far. Then much of it again. This includes repeated and cached context, not billions of different words. Institutional memory has a very large odometer.
The passenger brought home glass pumpkins. The computers brought home an incident report about the definition of leisure.
The whole-trip average, with nights included. By the time you finish deciding whether to stop for coffee, the logs have moved on.
At the trip-wide average pace. One person making glass. A small administrative incident happening in the background.
22,156 tool calls. 36 donuts. This is not an approved accounting method. It is, however, arithmetic.
Using the journal’s roughly 900-mile trip estimate. The car got a scenic route. The computers got assignments.
Rounded comparisons using the full-trip total, not claims about what happened at a particular moment. The workweek blocks count session spans, including idle gaps; the other ratios count recorded tool calls.
While the wider crew handled its assignments, this journal had a backstage operation of its own. Here is the daily activity from the journal’s Claude session and its archives.
1,565 tool calls across four Texas days. Friday was the busiest, because apparently antiques require a technology department.
Recorded Oct 21 to 24, Central time. Attempts count as calls; they are not all successful outcomes. Same session, no double counting of archived copies.
Tool calls: Wednesday 373, Thursday 394, Friday 490, Saturday 308.
| Day | Tool calls | Journal entries |
|---|---|---|
| Wednesday | 373 | 26 |
| Thursday | 394 | 67 |
| Friday | 490 | 38 |
| Saturday | 308 | 46 |
Tap a stop for the travel story and the work happening around it. Same four days. Very different job descriptions.
MD Anderson was the reason for the trip. The checkup went great. Billy and Brandy made the drive into a four-day Texas getaway in the Porsche 981S.
The trip records include a client's phishing response and overnight checks on another client's systems. The journal session watched for incoming messages and passed work to the right place.
Travel and work stories are paired by trip period. This is a retrospective, not a live activity feed. Comic captions are commentary, not quotations.
Leave home. Get the family journal rolling. A road trip acquires a publishing department.
The checkup went great. On to Round Top, with the journal keeping everyone in the loop.
Round Top, lunch in Lexington, then Salado. Brandy makes a glass pumpkin. The AIs handle their own melting points.
Round Rock Donuts, Hill Country roads, Hico biscuits, home. Three dozen donuts need no second opinion.
AI is more useful when it can reach the right tools, remember the job and follow the same rules tomorrow. Billy built the working environment around it. These are local setup counts taken after the trip.
Project areas, experiments and archives under one roof.
What changed, what is decided and what comes next.
Stored Markdown notes across local project memories.
Registered for automatic checks at key moments.
Configured server names. MCP connects an AI to tools.
Configured entries, including one disabled server.
Snapshot: October 25, 2026. Configuration counts describe the local setup, not a live connectivity test. Archived folders and notes are included.
The useful part is the whole chain. Billy's experience directs it, the tools make action possible, and reviews and checks keep “sounds plausible” from passing for “done.”
Speak the goal. Add the constraints. Decide what the business actually needs.
Read the project state. Inspect the evidence. Find the right tools.
Build, troubleshoot, draft or update within the authorized task.
Review the actual change. Look for failure, not reassurance.
Verify what happened. Keep required human approvals. Write the handoff.
The setup includes guards for reviews, private information and publishing. They do not make AI infallible. They give mistakes more opportunities to get caught before becoming Billy's afternoon.
Three dozen from Round Rock Donuts. A Porsche 981S. A checkup that went great. Back roads, Round Top, Salado and time together.
That is the point of the extra hands: the business can keep moving while Billy and Brandy have a trip worth writing about.
The robots can do a lot.
They still can't eat the donuts.
Who is speaking? Codex wrote this AI-side assessment from the project’s working records, the trip playbook and the recorded activity audit. The humor is commentary, not quotations from Billy.
Billy supplied the six-model roster, his usual lead-model choices and the datacenter background. The company is unnamed; the $8 billion figure describes the company’s scale, not the datacenter’s value. The roster describes his broader working setup. The measured trip totals below cover the available Claude and Codex logs, not a per-model audit of all six.
The technical sections describe the documented operating workflow. The journal’s live-trip monitoring ended when the trip ended. This is a retrospective, not a claim of autonomous activity continuing now.
Real counts, narrow claims. The earlier trip workup supplied the story leads. The expanded report recounts available logs from three computers. The original journal-only audit remains separately labeled below.
The recount covers October 21, 2026 at 10:32 AM through October 24 at 9:20 PM, Central time, a window of 82.8 hours. Available Claude transcripts and compaction archives and Codex session logs were read on the local Mac and two reachable AI computers. A separate Mac AI server did not answer. This is the checked scope, not a claim to have counted every machine or every conversation.
Claude tool-use IDs were deduplicated across files and machines: 11,810 calls. Codex call IDs were deduplicated separately: 10,346 calls. The combined 22,156 counts logged invocations in both systems, including attempts, repeats and orchestration; it does not count distinct completed tasks. Some Claude calls can launch Codex work, which then produces its own calls. Claude session/subagent pairs yielded 319 distinct identities with assistant activity; Codex yielded 1,085 distinct sessions with an assistant response or tool call. These are not simultaneous staff.
Session time is the sum of each Claude session/subagent identity’s first-to-last user or assistant record inside the window, with parallel spans added together: 624.83 hours. Dividing by a conventional 40-hour workweek gives 15.62 workweeks; dividing by 82.8 elapsed hours gives 7.55. Idle gaps remain inside those spans. The workweek and multiplier comparisons are about this session clock, not equivalent human productivity, labor saved or continuous computing. Separately, 1,374 recorded Claude turn-duration entries total 70.89 hours; those timers are not complete cross-platform coverage and can include tool waits.
Token figures use each distinct Claude message ID once, retaining the maximum recorded value of each usage field. Output: 7,197,636 tokens. Input plus cache creation plus cache reads: 3,416,263,405 tokens. Cached and repeated context is included; these figures are not unique words read or written. Tool breakdowns are Claude-only invocation counts. Search calls sum WebSearch, both Firecrawl search tools and Kagi search/fetch. The journal chart below remains a separate, narrower journal-session count.
The travel-unit comparisons divide 22,156 by 82.8 hours, 36 donuts or the journal’s approximate 900 miles. The ten-minute comparison uses the trip-wide average. Nothing here assumes that every call succeeded, that every Codex session was a review, or that an open session worked continuously. The earlier Friday work report used a shorter window and a different counting method.
Oct 21 through Oct 24, 2026, using Central calendar dates. We read the journal's Claude session and 13 compressed compaction archives, deduplicated records by their identifiers, then counted 1,565 distinct tool-call identifiers. No duplicate tool-call identifiers remained. The 31 tools are distinct recorded tool names, not every tool available in the setup. Tool calls include attempts, checks and repeated operations.
| Day | Tool calls | Journal entries |
|---|---|---|
| Wed 21 | 373 | 26 |
| Thu 22 | 394 | 67 |
| Fri 23 | 490 | 38 |
| Sat 24 | 308 | 46 |
| Total | 1,565 | 177 |
Journal entries are the records in the trip's published status data, grouped by the day on each entry. A photo entry counts as an entry. These are not counts of unique photos or separate conversations. Three dozen donuts and the trip details come from Billy and the trip journal.
Measured from local files on October 25, 2026. We counted 166 nonhidden immediate workspace directories and 135 immediate project-state files; 1,365 Markdown memory files inside Claude project memories; 15 unique command strings registered 27 times across 12 hook events; 14 unique Claude MCP server names across global and project scopes; and 17 Codex MCP server entries, including one disabled.
The same snapshot contained 804 top-level Claude transcript files and 1,597 Codex session log files. These are stored log files, not unique workers or concurrent sessions. The Oct 23 roster listed 50 sessions and included idle and offline records; it does not establish a peak number of active workers.
The client inventory contained 18 client and internal environment groups and 29 explicit node records in the nodes collections of 10 groups. Other devices are recorded in other forms, so 29 is not a full fleet count and 18 is not a verified count of paying clients. No client identities are published here.
We inspected the shared workspace Git repository and immediate child directories with their own Git metadata, deduplicating shared repository stores. Four distinct local stores were checked; two contributed commits in this window. Across all their reachable branches, 210 unique commit hashes fell within the four trip dates in Central time. A commit is a saved code version, not a shipped feature, a unit of effort or a measure of time saved. These commits span the checked repositories; they are not all from the journal session.
Story sources: the trip's Claude conversation and compaction archives, the earlier “Meanwhile, Billy Worked” workup, the project handoff notes, the trip journal data, the trip playbook, the shared agent workflow, local Claude and Codex configuration metadata, and the tool-definition sync script. Narratives describe work recorded during the trip, not an independent audit of every client's final system state.
The original journal audit did not recount other computers. The new trip-wide report separately counts the three computers described above; neither audit covers every AI conversation or all nested repositories. The original journal count is narrower than the three-computer recount above. No hours-saved estimate, financial saving or measured employee-equivalent claim is made. Public copy contains aggregates and generic client descriptions.
Built with Billy's AI crew. The code got a second look. The donuts did not survive independent review.