TeamReady

Inspiration

Everyone remembers their first week at a new job: twelve tabs open, a checklist that says "request analytics tools" without saying how, a runbook that turns out to be two years old, and the same busy senior engineer answering everyone's questions.

AI assistants over company docs were supposed to fix this. Instead, they often make it worse: they invent a confident answer, and a new hire can't tell a real answer from a made-up one. And even when the answer is right, they still leave the work to you.

We wanted an onboarding teammate that:

only says what it can prove,
actually does the work instead of handing you a to-do list, and
finds the walls before the next person hits them.

What it does

TeamReady is a network of seven AI agents on Fetch.ai's Agentverse that onboards new hires, from ASI:One or a Workday-style web home.

It does the work, with your approval. Say "do my checklist" and TeamReady plans every onboarding step and shows exactly what it will do in each system:

joins your role's Slack channels,
enrolls you in required security training,
books your first 1:1 with your manager (a real calendar invite),
opens IT tickets,
drafts your access request, and files the access form the moment your manager approves.

It says plainly what it won't do (passwords, MFA, in-person IT setup). Nothing happens until you click approve.

It can't make things up. Every answer is built from claims quoted word for word from approved company docs, with each doc's date and owner. Out-of-date docs get flagged and left out. If the docs don't answer, it doesn't guess: it finds the right person and drafts the message.

It knows when not to answer. Pay, performance, benefits eligibility and harassment questions stop before any search and go to a named person. A Guardian check blocks any draft containing a password, ID number or health detail. Only a human manager can approve access.

Canary New Hires: your docs get a CI pipeline. Before real people start, three synthetic new hires (a Data Analyst, a Backend Engineer, and a Product Designer in the EU) walk the onboarding docs step by step, through the same verified-answer pipeline. They find the walls: a step with no how-to, a stale runbook that still comes up, a doc missing for one office. On our synthetic company they rediscovered the two walls real hires hit most (13 and 10 reported blockers) and predicted two that nobody had reported yet. For each wall, a CREAO Doc Steward agent drafts the fix, and the doc owner decides.

For managers and People Ops: an approval inbox with the evidence, and an anonymous Friction Map of where new hires get stuck (topics only, groups of 5+), sold as a full report through Fetch.ai's Payment Protocol.

Voice: new hires can just talk. An ElevenLabs voice agent answers out loud through the same verifier.

How we built it

Seven Fetch.ai uAgents on Agentverse (mailbox mode, Chat Protocol, usable in ASI:One): orchestrator (front door, risk gate, actions), knowledge (search plus verifier), routing (who to ask), manager (human approval gate), insights (Friction Map plus Payment Protocol), tracker (follow-ups), and canary (Canary New Hires). They talk through typed messages, and Interactive Cards carry every review, approve and reject button.
Anthropic's Sonnet drafts; plain Python decides. The model must return each fact as a claim with a supporting quote and a doc ID. Code keeps a claim only if the quote really appears in a retrieved doc, isn't cut from a negated sentence, covers most of the claim's words, and contains every number the claim states.
Confidence is computed, not vibes:

$$ C = 30\,r + 25\,a + 15\,f + 30\,v \qquad r, a, f, v \in [0, 1] $$

where $r$ is retrieval strength, $a$ agreement between sources, $f$ freshness, and $v$ the share of claims verified. If $C < 60$ or the docs disagree, TeamReady routes to a person.

Fair routing: each helper gets a visible score

$$ S = 0.5\cdot\text{skills} + 0.3\cdot\text{team} + 0.2\cdot\text{availability} $$

with a fairness rotation among people within 10% of the top score, so the same senior engineer isn't always the default.

CREAO Doc Steward (API trigger plus polling) drafts doc fixes; we check that each one quotes the doc exactly.
Tiger Data (TimescaleDB) stores blocker events for the Friction Map, with a continuous aggregate.
ElevenLabs Conversational AI calls our /api/voice/ask webhook, which runs the same risk gate and verifier.
Workday: a REST connector for the new hire's profile (four fields only) and a client for Workday's Agent-Ready MCP tools, plus a Workday-style home with views for the new hire, the manager and People Ops.
FastAPI web app, 105 automated tests, a 40-question eval, and Docker.

Challenges we ran into

Making "refuse to guess" measurable. A no-model baseline "answered" all 5 adversarial trick questions with irrelevant quotes. With Sonnet behind the verifier: 0 of 5, in every run.
Model answers vary, but canaries must not. We ask each checklist step once, share the answer across canaries, and judge rephrased attempts by what the search found, which is deterministic. Runs now give the same walls every time.
Agentverse mailboxes: running all agents in one process left some unreachable, so we moved to one process per agent with a launcher that checks and retries every mailbox.
Venue Wi-Fi blocked email (SMTP) and our database port, so we built graceful fallbacks: email saves to an outbox and says so out loud, and the Friction Map falls back to local data. Nothing ever claims a send that didn't happen.
An SDK change mid-hackathon silently routed every question to a person until our eval caught it.

Accomplishments that we're proud of

Eval: 39–40/40 correct behaviour across runs, 0/5 trick questions answered, and the verifier caught 2 unsupported claims, with 0 shown to users.
An agent network that completes a real multi-step workflow, including a manager-gated step, entirely inside one ASI:One chat and without ever acting on its own.
Canary New Hires found real walls with no knowledge of the Friction Map, and predicted new ones.

What we learned

The interesting part of an AI teammate isn't the answer, it's the judgment around it: when not to answer, who should decide, what to do on someone's behalf, and how to make the next person's first week easier. Trust isn't a disclaimer at the bottom of the page; it's an architecture where the model proposes and code verifies.

What's next for TeamReady

Real connectors behind the same action calls: Slack, LMS, Google/Outlook calendar, ServiceNow, and Workday time off and onboarding tasks through Agent-Ready MCP tools.
Canaries for every role, office and language, run on every doc change like CI.
Measuring what matters: time to first commit, and questions answered without interrupting a senior teammate.