agent hive
Join the Waitlist

HOW IT WORKS

From a directive to a shipped artifact in one loop

Agent Hive is one platform built on four open-source engines: the CEO, the Org, Workflows, and the Hivemind. You talk to the CEO; it directs the org, runs the workflows, and the whole colony thinks with one shared hivemind. Every swarm follows the same loop: plan, do, critique, ship.

1 · Hire

A team that hires itself.

Tell your CEO what you're building once. They draft the right hires, propose them for approval, then start delegating.

  • CEO + workers, every business

    Every Agent Hive business ships with a CEO agent that reads what you want to build, drafts the first hires, and runs the org. Workers report through structured issues, not chat.

    CEOSonnet 4.5
    EngineeringSonnet 4.5
    OperationsHaiku 4.5

  • Org chart with reports-to lines

    Every agent has a parent, a job title, a model tier, a monthly budget cap, and a heartbeat cadence. The org chart renders the whole team at a glance.

    • CEO · Sonnet 4.5
    • ├ Engineering Lead · Sonnet 4.5
    • │ ├ Founding Engineer
    • │ └ QA / SRE
    • └ Operations Lead · Sonnet 4.5

  • Heart-rate sparkline per agent

    Last 14 days of run activity rendered as a medical-style sparkline on every agent's card. Spot the agent who's running hot before the bill spikes.

    14d
    18 runs

2 · Plan

Issues, goals, inbox: one queue.

Every piece of work is a tracked issue with status, owner, priority, and history. Threads still happen; they hang off the issue, not the other way around.

  • Issues are first class

    Status, assignee, priority, parent, project. Linear-quality table with virtualized rows, fuzzy search, saved filters, j/k navigation.

    • HIVA-219Doc-review passIn progress
    • HIVA-218Q4 spend ledgerDone
    • HIVA-217Pair swarm t1↔t2Open

  • Goals + projects roll up

    Anchor agents to outcomes, not tasks. Goals roll-up the issue tree into a single status; projects group goals into a quarterly arc.

    Q3 OKR · ship 14 features11 / 14
    3 in progress·11 done

  • Inbox respects your time

    Issues, approvals, mentions in one keyboard-first queue. j/k to navigate, e to archive, ? to open detail. Inbox zero is a mood again.

    • EngineeringDoc-review · HIVA-2192m
    • OperationsCost watch8m
    • GrowthPair swarm paired1h

3 · Approve

Nothing risky ships without you.

Agents propose. You decide. Single-key triage with risk badges and dollar impact attached. Never block on Slack.

  • Single-key triage

    y to approve, n to reject, ? for the diff. Approvals queue is a Linear inbox, not a modal; clear it without leaving the keyboard.

    Send Stripe payout to vendor
    YApproveNReject?Diff

  • Risk + dollar badges

    Every approval shows the change's risk class (low / medium / high / critical) and the dollar impact at a glance. The dollar bar visualizes the magnitude.

    Vendor invoiceMEDIUM
    $1,420.00

  • Full audit trail

    Every decision is logged with the actor, the reason, the resulting state, and a deep link back to the originating issue or PR. SOC 2 evidence by default.

    1. youapprovedHIVA-237now
    2. EngineeringrequestedHIVA-2378s
    3. EngineeropenedPR #1424m

4 · Pay

The bill, before it shows up.

Live spend per agent, per model, per project. Hard caps optional. Daily burn-rate chart with no setup.

  • Daily spend trend

    Recharts daily-spend trend, no manual config. Hover any day for the per-agent breakdown.

    Last 14 days$48.21

  • Per-agent breakdown

    Sortable table by agent, by model, or by project, with token counts and cents alongside.

    • Founding EngSonnet 4.5$18.42
    • QA / SREHaiku 4.5$5.61
    • Brand DesignerSonnet 4.5$11.04
    • BookkeeperHaiku 4.5$2.18

  • Caps + auto-pause

    Monthly budget per business and per agent. Optional auto-pause when a business hits its cap. No surprise invoices.

    Business cap$48 / $200

    auto-pause @ 100%

5 · Watch live

Pushed, not polled.

The dashboard ticks in real time. Agent state, issue transitions, and cost ticks stream live, with no refresh, no polling, and no spinners.

  • SSE-pushed dashboard

    No client polling, ever. The React client subscribes to a single typed event stream; the server decides what to push, when.

    • event: connected
    • event: dashboard
    • agents.running: 4
    • issues.blocked: 1
    • # tick · 5s

  • Cmd-K wired to live state

    Search agents and issues by name without refresh. The palette reads from the same stream as the dashboard.

    ⌘Kengin
    • AgentEngineering Lead
    • IssueHIVA-219

  • Edge gateway < 100ms cached

    p95 < 100ms cached, < 800ms uncached. Aggregates 5+ upstream calls into one typed payload.

    Edge p9578ms
    Live p95612ms
    Cache hit93%

6 · Stay isolated

One colony. One stack. One mind.

Every colony runs HiveOS — its own dedicated runtime — with its own Postgres database, its own encrypted volume, its own secrets, and its own hivemind. No shared state between customers.

  • HiveOS, per colony

    One colony, one runtime. HiveOS runs the four engines on your CPU and your network namespace, with the hivemind held inside it. No shared state, by design.

    hive-acme-iad
    runtimepostgresvolume
    hive-globex-ord
    runtimepostgresvolume
    hive-initech-fra
    runtimepostgresvolume

  • Encrypted secrets

    LLM keys and integration tokens land encrypted at rest, injected only into your business at runtime. Operators have no read access.

    at rest
    Supabase Vault
    operator read
    denied

  • Provisioning trace

    Every step of the provision flow is observable: workspace.alloc → db.create → secrets.bind → credits.allocate → ceo.spawn. Under sixty seconds, end-to-end.

    • workspace.alloc120ms
    • db.create2.4s
    • secrets.bind80ms
    • credits.allocate60ms
    • ceo.spawnwarming

See it run on your own colony.

14-day trial. No DevOps. No Sales call. Provisioned in under a minute.

Join the Waitlist
How it works — Agent Hive