We all use agents, right? NetworkChuck runs eight of them in his studio: Hermes, Codex, Claude Code and Pi-based local models. As Dodo puts it, the problem is no longer the model itself but managing the crowd. Who talks to whom, who does what, how is work tracked and where does security sit? Earlier attempts that just put agents in a room to chat never scaled. Paperclip enters here. Not a harness, a meta harness. Whether you run Claude Code, Codex, Hermes or Pi, the idea is to hire them all into the same company and make them talk through work. Presented as the clearest answer to harness fatigue, this video tests that answer live against a real outage.
Paperclip's creator Dodo shows his face on camera for the first time and sums up the meta harness in one line: agents become employees, you give them managers and you assign work, and communication happens through tasks, purposeful and precise. Instead of an aimless agent town, there are intentional tasks. The vision sold to the viewer is simple: when a new model or harness appears you do not rebuild from scratch, you add it to Paperclip and the existing company can use its new capabilities immediately. That is why harness maxing is repeated throughout. The company is framed as the single roof that can carry every latest model and harness feature, and coffee-break jokes keep that excitement alive.
The field mystery is very concrete: whenever someone flushes the toilet in the NetworkChuck studio, everyone loses the NAS connection. Editors included. The pattern is so regular the team oscillates between joke and seriousness. NetworkChuck turns the oddity into the long-held idea he wanted to try: build an IT department made of agents running on Paperclip, covering networking, storage and security. Acting like a real company, this structure will investigate step by step and log every step. Viewers are told to keep their own copy ready, because the narrative is staged as a build-along experience rather than a tutorial.
Where to run Paperclip splits into two paths. If you only want to try it, your current machine is enough: Mac, Linux or Windows with WSL 2 is the fastest start. For full use, a Proxmox VM or a cloud VPS is recommended. NetworkChuck promises to show both. In the Ubuntu VM example setup moves in two commands: first download the installer script, then run it. In advanced setup local network access is verified one by one, then onboarding starts in the browser at localhost:3100 or your domain. While creating the company and first account the CEO is named Dumbledore and the plan is to run that role with Claude Code.
The core team is assembled with a familiar naming language. Three Hermes agents named Ron, Fred and George form the IT backbone. Ron is CTO, Fred network engineer, George storage engineer. These agents are not trained from scratch; they are described as already existing, living in their own VMs and ready with the needed skills. NetworkChuck jokes it is like hiring from LinkedIn. The hierarchy is clear: Fred and George report to Ron, Ron reports to Dumbledore. On top, a Codex-based security engineer and a Pi-based local helpdesk are added. The message is that no matter which harness you use, Paperclip can gather them into one company and keep the company current with every new model.
The external threat surface is covered in the sponsored but instructive segment. Identity-focused threat intelligence platform Flare monitors the dark web, Telegram and hacker forums for exposed credentials, access and secret files. When NetworkChuck feeds his own data into Flare the picture is heavy: 76 exposed credentials, a ransom leak record, 390 messages, 23 lookalike domains and 32 thousand affected devices. These devices belong to real users of his academy and the leaks are for sale on a Russian market on Telegram at 10 dollars per log. Finding 300 thousand stored passwords on a single machine, stolen cookie sessions that bypass multi-factor authentication and attackers logging in instead of breaking in are highlighted. Infostealer malware harvesting passwords and sessions, 54 percent of ransomware victims showing leak traces long before the incident, companies detecting a breach in 36 hours on average while attackers monetize in 48 minutes, and Flare's Entra ID validation with automatic session revocation are tied together. NetworkChuck explicitly says he plans to connect this scan to Paperclip later via API.
The hands-on setup is shown step by step. On the Ubuntu VM two commands download and run the installer, quick start is recommended for most users while advanced setup verifies local network access one by one. The browser opens at port 3100, company onboarding starts and Claude Code is chosen for the Dumbledore CEO role. The first try fails because Claude Code is not yet installed and a failed inbox item appears. The fix is simple: install Claude Code in that box via copy-paste, sign in, return to Paperclip and retry. The dashboard onboarding continues, Dumbledore delivers a welcome message and the first company tasks are defined through the tasks tab. This section is crucial for clarifying that the meta harness itself is not an agent harness; the agent harnesses run on the Paperclip server.
The first hires make management and approval visible. After Dumbledore's onboarding task, Ron is invited. NetworkChuck messages Ron via Hermes and an approval request appears on screen. With approval Ron becomes an employee and says he is waiting for board approval. The camera turns to the org chart: the viewer as board at the top, agents as employees below. What is described as Paperclip's strongest side is demonstrated right here: communication is built not through chat but through tasks. For a test Dumbledore is given a check Ron task, Dumbledore instantly opens a sub-task to Ron, two agents work in parallel and when Ron finishes Dumbledore reports test complete. This is highlighted as the first moment one agent talked to another agent through a task.
The bulk hiring completes in a quick montage. Fred and George are added in one move, then Dumbledore hires the other roles through tasks. The resulting chart is: George and Fred on Hermes, Arthur Weasley on Codex, Watchdog on Pi, Helpdesk on Pi, Mad-Eye Moody for security review on Codex and Scanner again on Pi. Pi's good fit with local models is specifically noted. Roles and models diversify and the org chart branches. An important detail is added: you can talk to agents without entering the Paperclip UI, because every agent has the Paperclip skill. NetworkChuck shows this by making Ron assign a task to Mad-Eye via Hermes. The fact that the activity feed and every step is traceable and searchable is presented as essential when solving a large outage.
A project is created for the toilet-triggered outage and all known details are handed to Dumbledore. Dumbledore uses the experimental Cases feature to create a case and distributes work among Arthur, George and Fred. The task detail screen shows subtree pause, cancel, hide, blocker and parent task, reviewer, approver, watcher and case links. The escalation logic is especially emphasized: when an agent is blocked it does not bother the user directly, it forwards to Dumbledore first, and if Dumbledore cannot solve it escalates to the board, the viewer. Scanner scans in the field, suffers timeouts but progresses, Fred finishes his part and a studio network map plus a Markdown report produced by logging into switches appear in the artifacts tab. Arthur sets up a harness to catch events and gives Moody a security review, branching the parallel investigation.
Decision points and deeper investigation flow together. Dumbledore wants to proceed with three questions: does the NAS disconnect happen at the exact same moment for everyone, should a controlled flush test be run, and should optics be replaced? NetworkChuck says his suspicion concentrates on optics but chooses to wait for the report. New findings appear on the activity feed and artifacts screen: the SFP on the Mac side logged six link down events in a day with no match in storage. This is interpreted as proof the issue is not storage. Dumbledore makes Fred examine spanning tree and MAC table flushes, gives Ron correlation tasks and checks whether all NAS clients disconnect at the same second. As the team advances on multiple fronts at once, the fact that every finding is stored as an artifact and remains searchable is emphasized again and again.
Routines are presented as one of Paperclip's most loved and powerful sides in a separate demo. A routine is defined as a scheduled task and triggers on a set schedule. NetworkChuck builds a simple example via voice dictation: every day check storage on the SEF server, look for health issues, report remaining space and if below threshold produce a decision for the board. The task is assigned to Fred, a trigger for every day at 10 in the morning is added and the routine is run immediately. Fred returns in seconds and reports 29 terabytes free on the SSD tier. Although this number sounds large, it is counted as below threshold so a decision is produced but postponed for now. This demo is intentionally kept simple to show routines can even automate daily stand-ups.
The interim report and final diagnosis arrive in two acts. Dumbledore's interim report is honest: it is not the NAS, not the router, not spanning tree and one bad port is not pulling the others down. Despite deep dives the trigger is not identified. Although correlation with the flush looks strong, a door opening, an HVAC compressor running or a chair bumping the patch field are also possible. The concrete finding logged is that four fiber runs in the flu powder drop together within two seconds and produce 36 maintenance events per week. After a while the official verdict sharpens: not the toilet but four marginal links that have been degraded for 15 months. On the MikroTik switch ports 3, 5, 13 and 19 have no margin and lose sync 56 times a day. Port 5 has disconnected 20,052 times since boot, the neighboring port only 123 times. More than 40 weird disconnects per day and seven times during work hours 2 to 4 ports dropping together in the same second are explained only by these four links being at the edge. The blame falls on cheap third-party SFPs from Amazon bought in different batches; the failure rate in this batch is 50 percent and all problematic ports belong to the same batch. The fix is replacing them with three official MikroTik SFPs.
The closing merges both technical and cultural layers. While swapping inside there is a brief outage and the toilet room joke seals the moment. NetworkChuck summarizes the feeling of founding a company, hiring employees, making them talk through tasks and gathering all artifacts in one place. Although his belief in the toilet remains, the narrative notes correlation is still very strong and the link to events around the toilet persists. Viewers are told to watch for one to two weeks and asked to subscribe, the idea of a deeper electromagnetic environment test is postponed for lack of time. Then Flare automation returns to the stage: the Flare API key and tenant ID are created as secrets and given to a new agent named Filch, Filch scans with the Opus 5 skill and distributes what it finds to Ron for password resets and to Moody for reviews. The organization being importable and exportable, the daily stand-up routine triggering at 5 pm, the ten agents including thumbnail agent Lina and the dialogue between Taylor and Hermione that footage backups are expiring and restream recordings will lapse are presented as live examples of Paperclip culture, and the video closes with a request for comments and stars.
Key moments
- Eight agents, one problem: management and security bottleneck
Harness fatigue meets meta harness
- Toilet flush drops NAS: the weird studio correlation
Every flush disconnects the whole team
- Paperclip setup: two commands, port 3100 and Dumbledore CEO
- Agents talk via tasks: Dumbledore opens sub-task to Ron
- Artifacts show network map and SFP logs
Mac SFP six drops a day, no match in storage
- Routine demo: daily 10am 29 TB report
- Final diagnosis: MikroTik ports 3,5,13,19 marginal SFPs
Port 5 20k drops vs neighbor 123
AI commentary
"What stands out to me is how Paperclip shines by gathering every new harness and model under one company roof; the real appeal is steering dozens of agents through tasks from a single control plane while keeping every trace searchable."
AI assessment
Steel-manned, the case against Paperclip is also its appeal. A meta harness adds a second layer of abstraction on top of a single plane. For a team tired of harness fatigue, gathering every tool into one company sounds lean, but in practice every new harness wired into Paperclip, every model identity and every routine that must remain searchable adds maintenance. A leaner alternative is to orchestrate directly on one mature harness and limit agent-to-agent communication to API calls. For small teams or individuals on a single model, Paperclip's org chart, approval flows and artifact tracking can feel like ceremony that steals time from the actual work.
Methodologically the investigation runs on a single studio, a single MikroTik switch and without controlled experiments. Physical triggers like a flush, a door slam or HVAC were not measured with simultaneous sensors, no electromagnetic interference test was run, and voltage, vibration or temperature correlation along the cable route was not logged systematically. Post-swap validation is also short: it is not shown whether disconnect counts fell to zero for weeks after the three official SFPs were installed, whether other SFPs from the same batch were also replaced, and whether logs were re-read by an independent observer. The honest interim report is valuable, but a durable verdict needs long-term measurement and blind testing.
Through a provenance lens both Paperclip itself and the Flare integration are promotional. Dodo's project reaches a wide audience for the first time via the NetworkChuck video and the Flare sponsorship strengthens the identity intelligence narrative. That does not make the story false, but it does require independent replication. The strongest check would be margin measurements on the same fibers with the same switch, a controlled swap with SFPs from different vendors, and a numeric comparison of weekly drop events. On the Flare side, scale numbers like 32 thousand devices and 76 exposed credentials tied to academy users are verifiable against external sources, and the promise of automatic Entra ID session revocation should be tested together with real policy impact.
Practically, Paperclip makes sense for teams juggling multiple harnesses and models, creators with a home lab, and anyone who wants task-based orchestration to be logged. Turning a chaotic crowd of agents into an org chart, a task tree and searchable artifacts speeds up concrete failures like networking and storage. For individuals content with one model and one harness, low-latency financial systems or tightly regulated organizations, a lighter setup with direct API orchestration is less risky. Those who want to start should try quick start on a local machine and move to a VPS only when work branches; that keeps both learning curve and cost under control and Paperclip's real power emerges together with culture tools like routines and importable organizations.
Sources
7 links; no other published story cites them. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
- @youtube.com YouTube — NetworkChuck: you need to try Paperclip RIGHT NOW!
- @paperclip.ing https://docs.paperclip.ing/guides/getting-started/installation.md
- @paperclip.ing https://docs.paperclip.ing/reference/concepts/cases-and-routines
- @flare.io https://flare.io/identity-exposure-management
- @flare.io https://flare.io/blog/infostealer-credentials-dark-web
- @mikrotik.com https://mikrotik.com/products/group/sfp-modules
- @microsoft.com https://learn.microsoft.com/en-us/entra/identity/monitoring-health/concept-risky-users
paperclip · meta harness · networkchuck · mikrotik · flare