Fourteen engineers, three pods, one Slack thread titled "anyone actually using the agent thing?" That is the state of most engineering orgs six weeks after rolling out Claude Code with a multi-agent setup: a couple of early adopters running sessions daily, a majority who sat through one lunch-and-learn and never opened a session again, and a manager who cannot tell if the problem is training, trust, or the tool itself. The pattern repeats because the root causes repeat. Someone on the team had a bad first experience with a generic chatbot, it hallucinated a config key during a migration script or rewrote a function signature nobody asked it to touch, and that one incident becomes the story the whole team tells for months. Nobody owns adoption day to day; there is no engineer whose job it is to answer "how do I use this for X" on a Tuesday afternoon. And the one training session that happened in week one is a fading memory by week four, because a 45-minute Zoom call does not survive contact with a sprint. Onboarding a team to AI agents is a program, not a kickoff meeting: it needs a diagnosis of what actually broke last time, a golden path documented well enough that new hires follow it without being told twice, and a way to measure whether adoption is real or just attendance.
Why a generic chatbot and an autocomplete plugin both fail at onboarding a team to AI agents
The instinct to fix a stalled rollout is to schedule another training session, maybe with a better slide deck. That instinct misdiagnoses the failure. The original bad experience, the hallucinated config key, the overconfident rewrite, happened because someone pointed a generalist chatbot at a task that needed scoping, ownership, and a specialist, and got none of those. ChatGPT and Claude.ai do not distinguish between "explain this error" and "migrate our auth model," they answer both with the same undifferentiated confidence. When an engineer's first real use of AI on the job is a generalist producing wrong code with no scoping step in between, the lesson they internalize is not "I asked it the wrong way," it is "this doesn't work here." That lesson spreads through a team faster than any onboarding doc can counteract it.
Cursor and GitHub Copilot solve a narrower problem than team onboarding requires. They are excellent at completing the next few lines inside a file you already have open. But autocomplete has nothing to say about which of a hundred possible agents an engineer should reach for on day one, whether the org should surface all of them or a curated eight, what the golden path looks like for invoking a security review before a PR, or how a manager would know six weeks later whether the team is actually using any of it. Autocomplete lives at the keystroke. Onboarding a team lives at the org level: rosters, playbooks, champions, and adoption metrics. Neither the generalist chatbot nor the editor plugin operates at that level, which is exactly why "just point them at ChatGPT" and "just install Copilot" both produce the same six-week outcome: two engineers still using it, twelve who quietly stopped.
Folk designs the path, Pave documents it, Apex curates what day one looks like
Rebuilding a stalled rollout is a people-systems problem before it is a tooling problem, which is why it starts with Folk, tonone's org design and onboarding agent. Folk does not write code and does not decide which agent handles a migration; it designs how a team adopts new ways of working, including AI agents, and it does that by auditing what is actually happening before prescribing what should happen next.
Diagnosing why the first rollout stalled
Before Folk writes a single onboarding step, the folk-recon skill audits the current state: who is actually running sessions, who tried once and stopped, whether there is an informal champion already doing the work of answering questions in Slack without a formal role, and what specifically went wrong in the incident that soured the skeptics. This matters because "training didn't take" and "one bad experience poisoned the well" require completely different fixes. A recon that surfaces "engineer X hit a hallucinated env var during last quarter's migration and now tells anyone who asks not to bother" produces a different onboarding plan than a recon that surfaces "nobody told the team which agent to use for what." Skipping this step is why most redo attempts fail exactly the same way the first attempt did.
Tonone's Folk audits org design, hiring, comp, onboarding, and performance systems before prescribing changes, adoption problems get diagnosed, not just re-trained.
A playbook with checkpoints, not a single kickoff session
Folk's folk-onboard skill produces a day-1-through-week-4 onboarding playbook: what gets provisioned on day one, what the first paired session looks like, what a new hire, or an existing engineer being re-onboarded after a stalled rollout, should be able to do unsupervised by day five, and what milestone gets checked at week two and week four. The output is a checklist with owners and dates attached, not a slide deck that gets presented once and archived. For a rollout that already failed once, the checkpoints matter more than the content: the reason the first attempt faded by week four is that nobody checked in at week two to catch the engineer who quietly stopped.
Curating day one instead of dropping in the full roster
Handing a skeptical engineer all one hundred tonone agents on day one is its own failure mode, it recreates the generalist-chatbot problem of "which of these do I even use for this." Apex's apex-profile skill scopes the roster down to what the team's actual work requires: a devtools team writing backend services and shipping infrastructure does not need the legal, sales, or content agents in its day-one view. A curated roster of six to ten agents, matched to the pods actually doing the work, is what a new or returning user opens instead of a wall of a hundred options.
Tonone's apex-profile skill scopes the agent roster for a project, installing a curated subset instead of the full one hundred agent bundle, the difference between a golden path and an overwhelming menu.
The golden path doc that survives past week one
Once the roster is curated and the onboarding checkpoints are set, Pave's pave-golden skill defines the golden path itself: the opinionated, documented way the team invokes agents for its most common tasks, scaffolding a new service, running a security pass before merge, scoping a project before writing code. The output is concrete: steps, the exact slash commands to run, and where in the repo the templates live, not a philosophy of "AI-assisted development" that nobody can act on. This is the document that answers the Tuesday-afternoon "how do I use this for X" question without needing a champion to be online, which matters because champions go on vacation and change teams.
Making agent use a team norm, and proving it happened
Two more pieces close the loop. Folk's folk-culture skill documents the team norm explicitly, when agent-assisted work is expected, how it gets reviewed, what "done" looks like when a security or QA pass was part of producing it, so the practice is written down rather than left to whoever remembers the training. And Apex's apex-stats skill answers the question every manager actually cares about six weeks in: not who attended the session, but who is actually running sessions, which agents get used, and which ones never got touched. That is the honest adoption number, pulled from session transcripts, not a self-reported survey.
How Priya rebuilt onboarding at a 14-engineer devtools startup
Priya runs engineering at Ferrous, a 14-person devtools startup split across three pods, platform, product, and data. Three months ago the team adopted Claude Code with a multi-agent setup. The rollout consisted of one 45-minute all-hands demo in week one and a pinned Slack message with a getting-started link. By week six, two of fourteen engineers were running sessions regularly. The rest had one of two experiences: they never got past the demo, or they tried it once on a real ticket, got a plausible-looking but wrong migration script from a generalist prompt with no scoping step in between, and quietly went back to writing code by hand. One of the two, a senior platform engineer named Dev, had become the unofficial champion, answering Slack questions in his spare time with no recognition and no time carved out for it.
Priya brought Folk in to run folk-recon first. The audit surfaced three specific things: Dev was doing champion work informally and burning out on it, the bad first experience belonged to exactly one incident, a hallucinated Terraform variable during a staging migration in month one, that had since been retold at least four times in Slack, and nobody on the product or data pods had ever been shown which of the hundred available agents mapped to their actual work. None of that was visible from attendance numbers alone.
From the recon, Folk produced a redesigned onboarding playbook via folk-onboard, Apex scoped a ten-agent starter roster via apex-profile, apex, spine, proof, warden, pave, relay, vigil, atlas, flux, and forge, the set that actually matches platform and product pod work, and Pave wrote the golden path doc via pave-golden covering the four tasks the team does weekly: scoping a feature with apex-plan, running a security pass before merge, scaffolding a new service off Pave's own template, and getting a status report for standup. Dev's champion work got formalized: two hours a week, on the calendar, with a rotation planned for month three so it does not become one person's permanent second job.
Ferrous Engineering, AI Agent Onboarding Redo
Recon: 2/14 active by week 6 under v1 rollout. 1 soured incident (Terraform var
hallucination, retold 4x). No agent-to-task mapping for product/data pods.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Day 1 Curated roster installed (10 of 100 agents): apex, spine, proof,
warden, pave, relay, vigil, atlas, flux, forge.
30-min paired session with Dev on one real backlog ticket, not a demo.
Day 2-4 Each engineer runs apex-plan on one ticket they already own.
Champion (Dev, 2h/wk on calendar) reviews scope output, not code.
Day 5 First PR shipped using the golden path: apex-plan -> implementation
-> security pass -> merge. Golden path doc linked in PR template.
Week 2 Checkpoint: apex-stats pulled for the team. Target 8/14 weekly active.
Anyone below threshold gets a second paired session, not a nudge email.
Week 4 folk-culture norm published: when agent-assisted work is expected,
how it's reviewed, what "done" means. Added to onboarding doc for
all new hires going forward, not just this cohort.
Week 6 Checkpoint: apex-stats target 11/14 (79%) weekly active, up from
2/14 (14%) under the original rollout.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Recommendation: keep the ten-agent roster through month three, re-run
apex-profile once product pod's work shifts toward the data migration.Six weeks after the redo, Ferrous hit eleven of fourteen weekly active, up from two of fourteen under the original rollout, and the number came from apex-stats reading actual session transcripts, not a survey where everyone says yes because they do not want to look like the one behind. Dev's champion time is now a scheduled two hours a week instead of an invisible tax on his actual job. And the golden path doc, not a training deck, is what a new hire at Ferrous opens on their first day now.
| Capability | Tonone | Generalist chatbot | Cursor / Copilot |
|---|---|---|---|
| Diagnoses why a rollout stalled before redesigning it | Yes, folk-recon audits who's using it, who stopped, and what incident soured trust | No, no memory of your org's rollout or its history | No, autocomplete has no concept of team adoption |
| Curates a day-one roster instead of a full menu | Yes, apex-profile scopes 100 agents down to what the team's work actually needs | No, one undifferentiated assistant for every task | No, single-purpose completion, nothing to curate |
| Produces a golden path doc for common tasks | Yes, pave-golden documents concrete steps, commands, and templates | No, answers in the moment, nothing persists as team doc | No, no documentation output at all |
| Onboarding playbook with checkpoints, not one session | Yes, folk-onboard sets day 1, day 5, week 2, and week 4 milestones | No, one-off answers, no ongoing structure | No, no onboarding surface |
| Measures adoption from actual usage | Yes, apex-stats reads session transcripts for real usage per agent | No, no usage tracking capability | No visibility into who's using it or how often |
| Documents the team norm around AI-assisted work | Yes, folk-culture writes down when and how agent use is expected | No, no organizational memory | No, purely an individual-developer tool |
Tonone's apex-stats skill reads local session transcripts to show which agents a team actually uses, the honest adoption number, not a self-reported survey.
If your last AI rollout stalled by week six, don't re-run the same training session. Run /folk-recon first to find out whether the problem was one bad incident, a missing champion, or a roster nobody could navigate, then let /folk-onboard and /pave-golden build the playbook and the doc that fix the actual cause.
Install and try
Tonone is free and MIT-licensed. Install it once and Folk, Apex, Pave, and the rest of the roster are available in your team's Claude Code sessions. You pay only for the Claude Code token usage during work, and apex-profile lets you curate exactly which agents show up for your team on day one.
1. Add to marketplace
2. Install Folk
Frequently asked questions
How do you onboard a team to AI agents without adoption stalling?+
Start with a diagnosis, not another training session. Tonone's Folk runs folk-recon to find out whether the last rollout stalled because of a soured incident, a missing champion, or an unclear roster, then builds a checkpointed onboarding playbook with folk-onboard, backed by a curated day-one roster from Apex's apex-profile and a documented golden path from Pave's pave-golden.
What causes AI agent rollouts to stall after a few weeks?+
Three recurring causes: one bad early experience with a generalist chatbot that gets retold and poisons trust, no internal champion with dedicated time to answer questions, and training that happened once and was never reinforced with checkpoints. Tonone's folk-recon skill audits which of these is actually the cause before prescribing a fix.
How many AI agents should a team see on day one?+
Fewer than the full catalog. Tonone's apex-profile skill curates a subset, often six to ten agents, matched to what the team's pods actually do, instead of exposing all one hundred agents and recreating the same overwhelm that undermines adoption.
What is a golden path for AI agent usage?+
A golden path is the opinionated, documented way a team invokes agents for its recurring tasks: scoping a feature, running a security pass, scaffolding a service. Tonone's pave-golden skill produces this as concrete steps and commands, not a philosophy document, so it answers questions after the champion has gone home.
How do you measure whether AI agent onboarding is actually working?+
Measure usage, not attendance. Tonone's apex-stats skill reads local session transcripts to show which agents a team is actually running and how often, which is the honest number compared to a self-reported survey after a training session.
How do you support an informal AI champion on the team?+
Formalize the role instead of letting it stay invisible. Folk's folk-onboard playbook and folk-culture norm-setting give the champion scheduled time on the calendar and a documented expectation, with a rotation plan so the work doesn't become one person's permanent unpaid second job.
Is Tonone free to use for team onboarding?+
Yes. Tonone is MIT-licensed and free. You pay only for Claude Code token usage during actual work. Apex's apex-profile skill lets a manager curate the exact roster a team sees before any usage begins.
Which tonone agents help with onboarding a team to AI agents?+
Folk owns the onboarding design (folk-recon, folk-onboard, folk-culture), Apex curates the roster and measures adoption (apex-profile, apex-stats), and Pave documents the golden path (pave-golden). The three work together rather than as separate tools.