<- all tokdocs
2026.10.08@isaacdyor

Isaac Dyor says the cloud filesystem for agents is missing infrastructure you have to build yourself; Cloudflare, AWS, Turso and Mesa all shipped one between November 2025 and August 2026

AI
Watch on TikTok

View on TikTok ->

The artifact is a 92.77-second vertical clip, 1080x1920 at 30 fps, in a QuickTime/MP4 container (mov,mp4,m4a,3gp,3g2,mj2) whose video stream is HEVC Main profile, level 150, tag hvc1, pixel format yuv420p, carrying 2,783 frames at 282,137 bps, with an HE-AACv2 audio stream at 44,100 Hz, stereo, 64,121 bps; ffprobe reports a container duration of 92.841995 s against a video-stream duration of 92.766667 s, a total size of 4,093,774 bytes and an overall bitrate of 352,751 bps, matching the yt-dlp format id bytevc1_1080p_352748-1 and the filesize field in metadata.json. The post timestamp is epoch 1791445934, which is 2026-10-08 07:52:14 UTC. The uploader handle is isaacdyor (uploader_id 7236064524410373166) with the display nickname "Isaac Dyor", and the audio is credited as "original sound - Isaac Dyor" with artists set to ['Isaac Dyor'], so there is no licensed music track. The description is the title repeated, "Goated filesystem.", 18 characters, which carries no information about the subject. At the capture time recorded by the metadata.json file mtime, 2026-10-19 00:50:58 local, the counters read 13,700 views, 456 likes, 63 comments, 114 reposts and 320 saves. I read 17 of the 46 extracted frames: 001, 004, 007, 010, 013, 016, 019, 023, 026, 029, 032, 036, 037, 038, 039, 043 and 046, which at the 2-second extraction interval correspond to 0, 6, 12, 18, 24, 30, 36, 44, 50, 56, 62, 70, 72, 74, 76, 84 and 90 seconds. The transcript runs 345 words across 14 segments. The frames show a motion-graphics explainer, not a demo. Frames 001 and 004 hold a fake news card headed "The Infra Report" with the kicker "AI · INFRASTRUCTURE", a red "BREAKING" chip reading "Updated 2 min ago", the headline "The big labs still haven't solved the file system", the standfirst "Every agent works on its own copy of your files, and no one else can see them." and the byline "By Isaac Dyor · Oct 8, 2026", under an overlay title "OpenAI and Anthropic haven't solved the file system"; frame 004 adds the OpenAI and Anthropic wordmarks below the card. Frame 007 is a full-bleed talking-head shot with the burned-in caption "Here's exactly how it works". Frame 010 is titled "01 · THE PROBLEM" with the right-hand label "every agent, its own copy" and three devices labeled "your computer", "teammate's" and a browser window, each holding a folder marked "copy A", "copy B" and "copy C". Frames 013 and 016 are titled "02 · TODAY'S FIX" with the right-hand label "git", a GitHub pill reading "your-repo" on branch "main", a "git clone" pill, and per-device diff chips reading "+12 -3", "+40 -7" and "+3 -19". Frame 019 is a title card reading "One folder. On every machine." around a folder glyph labeled workspace/. Frames 023, 026, 032, 037 and 038 are titled "03 · THE FIX" with the right-hand label "a cloud file system", a cloud holding workspace/, a "mount" pill, and three endpoints labeled "your laptop", "teammate's" and "agent sandbox", each listing footage/, script.md and edit.tsx; frame 032 highlights script.md v4 on the teammate node with a pencil badge. Frame 029 shows a media player card titled launch.mov with a partially hydrated checkerboard of video tiles over a Golden Gate Bridge shot, a scrubber reading 1:07:22, the marker "▲ downloaded" and a total duration of 2:14:25. Frames 036 and 037 are titled "04 · HOW WE BUILT IT" with an orange Cloudflare mark, a card headed "Durable Object" annotated "1 per workspace" and "tiny server + SQLite · knows every file", listing footage/launch.mov v12, script.md v4 and edit.tsx v31, a broadcast pill reading "change #1042 · script.md v4 → all", and a second card headed "R2" annotated "the actual bytes" holding nine truncated content hashes (9f2a…, 41be…, c07d…, e81f…, 3a9c…, 77d0…, b52e…, 0c4f…, d9a1…). Frames 039 and 043 are titled "05 · WHAT IT UNLOCKS", showing "you" and "teammate" plus Anthropic-marked and OpenAI-marked "agent" nodes around a shared cloud with a "✓ live collaboration" pill, then three numbered browser windows under an "all the state" arrow pointing up into the cloud. Frame 046 is a white abstract logo glyph on black with no accompanying wordmark, so the product name never appears on screen in any frame I read, and no terminal, editor, file manager, repository URL, version number or running instance of the system appears in any frame I read. The spoken argument is that every agent today works on its own copy of the files, that Git is the current workaround and is neither live nor suitable for video, and that the fix is a cloud filesystem mounted on every machine with a Cloudflare Durable Object tracking file state and R2 holding the bytes, which in turn lets sandboxes become ephemeral. The transcript-end check passes: the last Whisper segment ends at 90.80 s against an audio.mp3 duration of 92.891429 s, so no segment is timestamped past the end of the audio and nothing in the transcript needs to be discarded as hallucinated.

The per-agent-copy premise is accurate, and Anthropic's and OpenAI's own documentation says so

At 13.18 s the video says "each agent operates on its own file system, whether that's your computer, your teammate's computer, or a cloud sandbox, which means that they can't see each other's work", and at 21.94 s it adds "The common solution to this is Git. Each agent clones your repo, works on its own copy, and then sends back its changes in a pull request." This is the part of the video that holds up without qualification, and the primary sources are the vendors' own docs.

Anthropic's Claude Code documentation describes exactly the architecture the video draws in frames 013 and 016. On cloud sessions it states that "each session runs in an isolated, Anthropic-managed VM", and on starting one from the terminal it states that "The cloud VM clones your current directory's GitHub remote at your current branch, not your local checkout, so push first if you have local commits." The handoff back is a pull request: "When a session completes, you can create a PR from claude.ai/code or teleport the session to your terminal to continue working." Pulling a session local does not merge the two views either. The same page says "The terminal gets its own copy of the session: new work there stays local and doesn't appear in the cloud session on claude.ai or the Claude mobile app." Even the sharing feature is a snapshot rather than a live view: "Recipients see the latest state when they open the link, but their view doesn't update in real time."

OpenAI's Codex cloud documentation says the same thing in one sentence: "Each task has its own workspace and can keep working while your computer is asleep." The environment is the reusable part, not the files. The docs frame it as "Reuse your repositories, dependencies, and tools across tasks. Each task has its own workspace."

So the diagram in frame 010, three folders labeled "copy A", "copy B" and "copy C", is a fair description of what both vendors ship. The video's problem statement is verified.

Correction: the infrastructure was not missing on 2026-10-08. Cloudflare shipped its own version two months earlier

At 0.00 s the video opens with "The file system is at the core of every AI workflow, and yet OpenAI and Anthropic are still missing the piece of infrastructure that allows you, your teammates, and your agents to all work on the same files", and at 41.08 s it tells the viewer "You basically need to build a file system that lives in the cloud, and then mount it on each machine as though it was a local folder." Taken together with "We had to build this for ourselves at Compound" at 9.86 s, the video's framing is that this did not exist and had to be built. That framing is wrong by a wide margin, and the clearest counterexample comes from the same vendor the video says it built on.

Cloudflare published "Your agent needs a computer, not a container — introducing @cloudflare/computer" on 2026-08-03, two months and five days before this video posted. The package is described as "an agent runtime that dynamically orchestrates between fast, efficient isolates and full Linux containers", providing "a durable filesystem that you can use with git repositories, storage buckets or any files you choose", with the workspace "backed by SQLite". That is the same two-layer design the video presents in frames 036 and 037 as something Compound had to build: a SQLite-backed durable workspace in a Durable Object, with object storage behind it.

Cloudflare shipped the mount half even earlier. Its changelog entry dated 2025-11-21 states that "Containers now support mounting R2 buckets as FUSE (Filesystem in Userspace) volumes, allowing applications to interact with R2 using standard filesystem operations", and the accompanying example documents the mechanism, using "tigrisfs", described in the docs as "an adapter for S3-compatible storage", to produce a mount at /mnt/r2. Cloudflare also maintains cloudflare/artifact-fs under Apache-2.0, a FUSE driver whose README says it is "designed to mount large git repos as quickly as possible, hydrating file contents on-the-fly instead of blocking on the initial clone" and that "the file tree is visible immediately after mount, and reads block only until the requested blob is fetched." That is the behaviour frame 029 illustrates with the half-hydrated launch.mov tiles.

Three more shipped products predate the video. Amazon announced on 2026-05-06 that Bedrock AgentCore Runtime supports attaching S3 and EFS access points to agent runtimes, stating that "AgentCore Runtime mounts the file system into every session at a path you specify, and your agent reads and writes files using standard file operations" and that "Multiple agents, or multiple sessions of the same agent, can collaborate on the same dataset." Mesa Systems, Inc. published "Introducing Mesa: a versioned filesystem for agents" on 2026-04-28, describing "a fully posix-compatible, durable filesystem with built-in version control", mountable with a FUSE daemon, with "sparse materialization so massive sets of documents load instantly and only files needed by the agent are fetched on demand". Turso maintains tursodatabase/agentfs under the MIT license, a SQLite-backed agent filesystem whose README states that "Every file operation, tool call, and state change is recorded in a SQLite database file."

The accurate version of the claim at 0.00 s is narrower: OpenAI and Anthropic do not ship a shared live filesystem in their own agent products, which their docs confirm, but the category of mountable cloud filesystems for agents was an active and shipped one for at least five months before this video posted.

The Cloudflare architecture the video describes is real, and the Durable Object has a 10 GB ceiling the video does not mention

At 67.06 s the video says "We built ours on Cloudflare. A durable object keeps track of every file and every change, and R2 stores the actual bytes of the files", which frame 036 renders as a card reading "Durable Object · 1 per workspace · tiny server + SQLite · knows every file".

The primitives exist as described. Cloudflare's documentation defines a Durable Object as "a special kind of Cloudflare Worker which uniquely combines compute with storage", each with a "globally-unique name" and "durable storage" that is "strongly consistent yet fast to access". Cloudflare's changelog entry "SQLite in Durable Objects GA with 10GB storage per object", dated 2025-04-07, took the SQLite backend to general availability, and a later changelog entry dated 2026-07-09 made it mandatory for new namespaces, so by the time this video posted a new Durable Object class would use SQLite by default.

The ceiling matters for the claim at 55.02 s, where the video says "you can work on hundreds of gigabytes of video without filling up your disk" over a frame 029 scrubber showing a 2:14:25 source file. The split architecture makes that possible, but only because the heavy half is R2, not the Durable Object. The SQLite database inside a Durable Object is capped at 10 GB per object, which is the number in the title of Cloudflare's own GA changelog. R2 is where the headroom is: Cloudflare's R2 limits page gives a maximum object size of "5 TiB per object", a single-part upload limit of 5 GiB, a multipart ceiling of 4.995 TiB across a maximum of 10,000 parts, and a maximum of 1,000,000 buckets per account. So "hundreds of gigabytes of video" is consistent with the design, provided the Durable Object holds only metadata and version pointers, which is what frame 036's three-row table of v12, v4 and v31 implies. The video never says the index has a size limit, and a viewer could reasonably come away thinking the Durable Object holds the workspace.

Verified: Git does fall apart on large files, and GitHub publishes the thresholds

At 28.30 s the video says "Agents only see changes when someone pushes and someone else pulls, and Git falls apart on large files like video." GitHub's own large-files documentation supplies the numbers. It states that "If you attempt to add or update a file that is larger than 50 MiB, you will receive a warning from Git", that "If you add a file to a repository via a browser, the file can be no larger than 25 MiB", and that "GitHub blocks files larger than 100 MiB." For anything above that, "To track files beyond this limit, you must use Git Large File Storage (Git LFS)." The same page notes that "Git is not designed to handle large SQL files" and recommends releases rather than tracked binaries for distributing large files.

The limit is not abstract for agent workflows. Anthropic's cloud-session docs impose a matching constraint on the bundle path: "The bundled repository must be under 100 MB. Larger repositories fall back to bundling only the current branch, then to a single squashed snapshot of the working tree, and fail if the snapshot is still too large." The same docs note that Claude Code "leaves out uncommitted changes to files that a git filter such as Git LFS manages", so the LFS escape hatch creates its own gap in what a cloud agent can see. A 2:14:25 video file of the kind frame 029 depicts is far outside the range Git handles without LFS, and the video's claim is correct.

Clarification: mounting object storage as a filesystem has costs Cloudflare documents and the video omits

At 47.42 s the video says "Your laptop, your teammate's laptop, and the agent's sandbox all see the exact same folder, but the bytes don't actually live on any of them", and at 55.02 s it adds "When you open a file, it streams just the parts that you need". Frames 023, 026 and 032 draw this as three identical file listings hanging off one cloud with a "mount" pill, and frame 032 shows the teammate node editing script.md v4 with the change propagating. Frame 043 then labels the state consolidation "all the state", which supports the claim at 80.64 s that "you can treat all of your sandboxes as ephemeral workers instead of needing them to be persistent machines."

Cloudflare's own FUSE documentation is more cautious than the video about what this buys you. Its R2 mount example states plainly that "FUSE provides filesystem access to object storage. It does not provide local-disk performance or full POSIX filesystem semantics", and recommends the pattern for "reading datasets, models, and shared assets". The Cloudflare changelog lists the intended uses as "bootstrapping containers with datasets, models, or dependencies", "persisting configuration or application state without managing downloads", and "accessing large static files without bloating container images or downloading at startup". The example in the docs mounts the bucket read-only. Mesa's launch post is the one source here that advertises both POSIX compatibility and concurrent agent writes, and it does so by adding version control and branching, describing "branches so agents can work in parallel without locking", which is a different concurrency model from the single live folder the video draws.

The video presents instant propagation at 62.28 s: "when anyone changes a file, it goes straight back to the cloud, and all the other computers see it instantly." Nothing in the frames or the transcript addresses write conflicts, locking, or what happens when two agents edit script.md at the same time, and the per-file version counters in frame 036 (v12, v4, v31) are the only hint that a versioning scheme exists. This is the largest unexamined gap in the argument.

Unverified: the frames never name the product, and I could not confirm what Compound is from a primary source

At 9.86 s the video says "We had to build this for ourselves at Compound, so I'll show you exactly how it works", and the on-screen text that follows in frame 007 reads "Here's exactly how it works". It does not show how it works. Across the 17 frames I read there is no terminal, no shell output, no file manager, no editor, no repository URL, no package name, no version string, and no running instance of the system. Frame 046, the final card, is an abstract white glyph with no wordmark beside it. Every technical frame is a diagram.

I could not establish what Compound is from a primary source. A web search returns a LinkedIn result whose title reads "Isaac Dyor - Co-Founder @ Compound", but linkedin.com returned HTTP 999 and I could not open the profile, so I do not cite it as established. His own site at isaac.dyor.com describes him as "a 20-year-old engineer" who works at Wordware "building agents" and lists no projects or posts, with no mention of Compound or a cloud filesystem. His GitHub profile at github.com/isaacdyor gives the bio "Software engineer interested in startups", the company field "The Corporation", a link back to isaac.dyor.com, and one pinned repository, t3-blog, with no mention of Compound. The company at getcompound.ai is a different business, described on its own site as an AI analyst for institutional finance serving private equity, venture capital, growth equity and hedge funds, and its named co-founder is Yao Zhao, not Isaac Dyor. I found no repository, changelog, documentation page, pricing page or launch post for the filesystem described in this video. The name "Isaac Dyor" is spelled consistently across his personal site, his GitHub profile and the byline rendered in frame 001, so the spelling is confirmed even though the company is not.

The architecture is sound and matches shipped Cloudflare primitives. The existence of a working implementation at a company called Compound is an assertion the video makes about itself and nothing I fetched corroborates.

Key Takeaways

  • Verified: The core problem statement holds. Anthropic's Claude Code docs say "each session runs in an isolated, Anthropic-managed VM" that "clones your current directory's GitHub remote", and OpenAI's Codex cloud docs say "Each task has its own workspace", which is precisely the "copy A / copy B / copy C" diagram in frame 010.
  • Correction: At 0.00 s the video says OpenAI and Anthropic "are still missing the piece of infrastructure" and at 41.08 s that "You basically need to build" it. Cloudflare published @cloudflare/computer, a SQLite-backed durable agent filesystem, on 2026-08-03, two months before this video; it shipped R2 FUSE mounts in Containers on 2025-11-21; AWS shipped mountable S3 and EFS filesystems for Bedrock AgentCore Runtime on 2026-05-06; Mesa published its versioned agent filesystem on 2026-04-28. The category was not missing.
  • Partial correction: At 55.02 s the video says you can "work on hundreds of gigabytes of video" on this architecture. That is true only because R2 holds the bytes, with a documented 5 TiB per-object ceiling. The Durable Object half is capped at 10 GB of SQLite per object, per Cloudflare's own GA changelog of 2025-04-07, and the video never mentions a limit.
  • Verified: At 28.30 s the claim that "Git falls apart on large files like video" matches GitHub's published thresholds: a warning above 50 MiB, a 25 MiB browser upload cap, a hard block above 100 MiB, and Git LFS required beyond that.
  • Clarification: The video presents the mount as transparent. Cloudflare's own R2 FUSE docs state that "FUSE provides filesystem access to object storage. It does not provide local-disk performance or full POSIX filesystem semantics", and the documented example mounts read-only. Concurrent writes, locking and conflict resolution are never addressed in the transcript or the frames.
  • Unverified: I could not confirm from any primary source what Compound is or that it ships the system described. Isaac Dyor's own site lists Wordware, his GitHub lists "The Corporation", getcompound.ai is an unrelated finance company, and linkedin.com returned HTTP 999 so I could not read the profile whose search-result title names Compound.
  • Context: Frame 007 burns in the caption "Here's exactly how it works", but across the 17 frames I read there is no terminal, no code, no repository and no running instance. Every technical frame is a motion-graphics diagram, and the final card at frame 046 shows a logo glyph with no wordmark.
  • Context: The transcript is clean. The last of 14 Whisper segments ends at 90.80 s against an audio.mp3 duration of 92.891429 s, so nothing in the transcript is timestamped past the end of the audio.

Resources

Published October 8, 2026. Writeup generated from a favorited TikTok.