Hive Research Institute AI Practicum
Block 2 · Agents, Not Chats

Hive Research Institute · AI Practicum · Block 2

Agents, Not Chats:
Fable 5.1 Max in the Enterprise

A live GrokBot narrative: what an agent does that a chat can't, how we keep it safe, and how it runs your week while you run the company.

Built by Claude Fable 5.1 Max This deck was researched, written, designed, coded and tested by the model in one agent session.

Press → or Space to begin · ? for shortcuts

Thesis

A chat answers. An agent finishes.

Chat

  1. 1You ask
  2. 2It answers
  3. 3You do the work
  4. 4It forgets

Value stops when the tab closes.

Agent

  1. 1You delegate an outcome
  2. 2It plans, uses tools, checks its work
  3. 3It reports, and asks only for decisions
  4. 4It remembers, and runs again tomorrow

Value accrues while you're elsewhere.

The unit of work moves from prompt to delegation.

Where Fable sits · connects to this morning's comparison

Three frontier models. Three different jobs.

Today's hero

Claude Fable 5.1 Max

Depth and judgment for long-horizon work

  • Runs multi-hour, multi-tool jobs without drifting
  • Careful with permissions, explicit about uncertainty
  • The model you hand a job to, not a question

GPT-6 Astra

Breadth and ecosystem reach

  • Generalist across consumer and enterprise surfaces
  • Widest plug-in and integration surface
  • Strong default for broad knowledge work

Gemini 3.8 Flash

Speed and cost at scale

  • High-volume, low-latency tasks
  • Bulk multimodal ingestion
  • Built for pipelines, not delegation

Positioning as used in this practicum, September 2026. Not a benchmark: all three move monthly. Pick per job, not per brand; most enterprises will run more than one.

Capability map

Five things an agent has that a chat doesn't.

Persistence

Memory that survives the session: your preferences, your org, last month's decisions.

"Same board-prep format as last month, but add the hiring plan."

Tools

Reads and writes in real systems: calendar, mail, CRM, Notion, Asana, files.

It doesn't describe the update. It makes it.

Routines

Scheduled and event-triggered runs: nightly, weekly, "when a deal moves stage".

06:45 brief in your Slack DM, every weekday.

Multi-app orchestration

One job that spans five systems, with the hand-offs between them handled.

Meeting → CRM → Notion → Asana → Slack, in one run.

Away-mode

Works while you're in meetings, on a flight, or asleep. Escalates only for decisions.

You come back to a decision list, not a to-do list.

Meet the operating layer

GrokBot is the operating layer. Fable is the mind inside it.

Interface Slack DM · chat · voice memo · forwarded email Where you delegate
GrokBot runtime Memory · routines · orchestrator · run log What persists
Reasoning Claude Fable 5.1 Max Plans, decides, writes, checks
Aegis gate Policy · 1Password vault · audit ledger Every credential passes here
Systems Calendar · Mail · CRM · Notion · Asana · Files Where the work lands

Swap the model, keep the layer. Swap a system, keep the agent.

Live demo · Beat 1

Delegate, don't prompt.

You · 07:02

Prep me for Thursday's leadership meeting: open decisions from Notion, CRM movement in our top accounts, and a draft Slack update. Interrupt me only for decisions.

GrokBot · plan
  1. Notion → open leadership decisions
  2. CRM → tier-1 changes, last 7 days
  3. Calendar → Thursday conflicts
  4. Slack → draft, hold for approval
  • NotionQuery "Decisions" · 7 open, 3 need leadershipdone
  • AegisCRM credential · read-only · tier-1 scope · 15 mingranted
  • CRM4 changes: 2 advanced, 1 stalled, 1 new ownerdone
  • CalendarThu 10:00–11:30 holds · 1 attendee conflict flaggeddone
  • SlackUpdate drafted · not postedawaiting you
GrokBot · 07:09

Ready. Three decisions need you: the vendor renewal (option A or B), the Q4 hiring count (2 or 3), and whether to move the offsite. The stalled deal has a follow-up drafted for its owner. Say send to post the update.

Live demo · Beat 2

It works while you're away.

You
In meetings, on the road, offline
GrokBot
05:30 Nightly routineInbox · CRM hygiene · Notion digest
06:45 Morning briefDay plan and decision queue, in your Slack DM
12:30 Midday checkCalendar changes · escalations only
17:30 WrapTomorrow's prep · follow-ups drafted
ScheduledCron and event triggers
TieredActs only inside its permission tier
EscalatesOne-tap approve in Slack
AuditableEvery run leaves a log
StopsUnsure? It pauses and asks

Use case · The same machinery at home

Family management: one inbox, many sources.

Inputs

  • School portal and class emails
  • Shared family calendar
  • Household list and renewals
  • Bookings, forms, appointments
Household agent

Separate memory · separate vault · no access to work systems

Outputs

  • Sunday-evening week brief
  • Reminders to the right person
  • Forms and deadlines tracked to done
  • Conflicts flagged before they bite
Sunday 18:00 · Week brief (illustrative)

Two pickups swap on Wednesday. One school form is due Thursday, pre-filled for signature. Friday's dentist slot clashes with practice; two alternatives proposed. Groceries list updated from the meal plan.

Details stay generic on purpose. Names, schools and addresses live in the household vault; the model works with references.

Security

The model never holds a secret.

Aegis is a small gatekeeper service between the agent and 1Password. The agent asks for capabilities; Aegis decides, fetches, scopes, executes and logs.

  1. Request"Read tier-1 deals in the CRM."
  2. PolicyAllowed for this agent, this tier, this hour?
  3. Fetch1Password service account → short-lived, least-privilege token
  4. ExecuteThe call runs in an isolated executor. The model sees results, not keys.
  5. LogWho, what, why, when → the audit ledger
What the model sees crm.read(deals, tier: 1, cred: vault://crm/readonly) A reference, never a value. Rotate the key in 1Password; nothing else changes.

Governance · What the agent may do

Write the trust ladder before the first run.

Tier 1

Acts alone

  • Read, summarise, brief
  • Draft anything
  • Update internal trackers
  • Schedule with colleagues
Tier 2

Proposes, you approve

  • External email or messages
  • CRM record changes
  • Publishing to shared spaces
  • Spend under a set threshold
Tier 3

Never without a human

  • Payments above the threshold
  • Deleting data
  • Contracts and commitments
  • Touching a raw secret

Aegis enforces the ladder in code. Good intentions are not a control.

Workflows that run today

CRM and Notion: the agent keeps them true.

CRM

pipeline hygiene
  1. After every meetingNotes to the record, next step and owner set, follow-up drafted
  2. WeeklyStale deals and missing next steps flagged to their owners
  3. Before meetingsA one-page account brief in your DM, 30 minutes prior

Notion

the company brain
  1. Decisions logKept current from meeting notes, with source links
  2. Weekly digest"What changed" for the whole company, every Friday
  3. OKR check-insPre-filled from Asana and CRM data; owners edit, not author

Guardrail. Writes land as proposals with a source link. Humans promote to canonical. Every record shows which agent touched it, and why.

Orchestration

One orchestrator. Many specialists. One gate.

  • OrchestratorFable 5.1 Max plans, delegates, checks, reports
  • SpecialistsNarrow job, narrow permissions: research, CRM, Notion, calendar, inbox, Asana
  • AegisEvery credential for every agent passes one gate
  • HumansOwn decisions and approvals; read the run log

Adding a specialist is a config change, not a project.

Drag the model to rotate

Next up

Mario Montag

Team dynamics and agent management

AsanaSlackGrokBot

You've seen one person delegate to one agent. Mario shows what happens when a team does it: agents with owners, tasks and threads, held to the same standards as people.

Takeaways

Run this tomorrow.

  1. Pick one weekly job you already do.Write it as a delegation with an outcome, not a prompt.
  2. Give the agent memory and two tools.Calendar and Notion are enough to start.
  3. Put every credential behind a gate.No secrets in prompts. Ever.
  4. Write the trust ladder.Three tiers, one page, signed by you.
  5. Read the decision inbox daily for two weeks.Then add the first routine.

Agents, not chats.

Over to Mario: Asana · Slack · GrokBot.