Hive-Mind Agent

One brain for your whole engineering team

Backend devs, frontend devs, and product managers ship the same product. With Headkey's hive-mind agent, their Claude Code sessions share one memory, one set of beliefs, and one knowledge graph — across every repo, without manual handoff.

FrontendWaits on backend to explain the API they just shipped.
ProductPings engineers for status instead of reading it directly.
EveryoneLast month's architecture decision resurfaces as a prod bug.

Scenario 01

Backend ships. Frontend knows.

The backend engineer captures the new endpoint’s contract and gotchas in their Claude Code session. Hours later, a frontend engineer in a different repo asks about the orders API — and gets the same answer, inline, without a Slack thread.

Backend — Jordansavantly/api-gateway
learn
{
  "content": "POST /v2/orders now uses cursor pagination (opaque string in `next`), not offset. Cursor expires after 15 minutes. Shipped in api-gateway PR #482.",
  "intent": "observation",
  "tags": [
    "api",
    "orders",
    "pagination"
  ],
  "importance": "high"
}

Response

Stored as memory. Linked to entities: Orders API, Pagination.
No conflicts with prior beliefs.
Frontend — Priyasavantly/web-app
ask
{
  "query": "how does the orders API pagination work?"
}

Response

POST /v2/orders uses cursor pagination — pass the opaque string from `next`
into the following request, not an offset. Cursors expire after 15 minutes.
Shipped in api-gateway PR #482.

Related: 3 memories, 1 belief ("cursor pagination is the team standard", 0.82)

Both engineers authenticate as the same team agent via OAuth. Writes land in the shared scope; reads pull from everything the team has learned.

Scenario 02

Monday morning, already caught up.

A single ask pulls memories, beliefs with confidence scores, and open decisions from every engineer’s prior week. No more scrolling Slack to rebuild context.

PM — AlexClaude Code (standalone)
ask
{
  "query": "what did the team learn and decide last week?",
  "timeRange": "last_7_days"
}

Response

See digest →

Weekly digest

Shipped
  • Cursor pagination on /v2/orders (api-gateway #482)
  • Rate-limit middleware on the public edge (api-gateway #491)
  • Dark-mode token refresh (web-app #1240)

Beliefs formed or reinforced
  • "Cursor pagination is the team standard" — 0.82, reinforced Tue, Thu
  • "Edge rate-limits must be IP+agent, not just IP" — 0.78, new Wed
  • "Dark-mode tokens need SSR-safe hydration" — 0.71, reinforced Fri

Decisions
  • Moved Stripe webhooks off direct DB writes, onto the outbox pattern (ADR-017)
  • Deprecated /v1/orders — sunset April 30

Open threads
  • Session-cookie migration still under discussion (2 contradicting beliefs, scorer
    surfaced a conflict on Thursday — no resolution logged yet)

Unlike flat RAG, the digest separates facts (memories) from opinions (beliefs with confidence) — so Alex can see what the team decided, not just what was said.

Scenario 03

Contradictions surface in-flow, not in incidents.

When a team member writes something that contradicts an existing belief, Headkey’s LLM scorer decides between reinforce, weaken, qualify, contradict, or create. High-confidence verdicts apply silently; low-confidence verdicts surface for a human call.

Backend — Jordansavantly/api-gateway
learn
{
  "content": "We switched auth from JWT bearer tokens to signed session cookies. Backend rolled this out yesterday in api-gateway PR #501.",
  "intent": "observation",
  "tags": [
    "auth",
    "architecture"
  ]
}

Response

Stored as memory. Scorer found 1 belief conflict below threshold:

  conflicts[0]:
    priorBeliefId:   blf_8f2a...
    priorStatement:  "Auth uses JWT bearer tokens" (confidence 0.88)
    newStatement:    "Auth uses signed session cookies"
    verdict:         contradict
    confidence:      0.64
    resolutionHint:  "Your new statement contradicts a prior belief —
                      is the prior outdated?"

  autoResolved: []

Claude Code will prompt you: apply the correction?

In Claude Code

Headkey found a conflict. Prior belief: “Auth uses JWT bearer tokens” (0.88). You're now saying: “Auth uses signed session cookies”. Scorer verdict: contradict at 0.64. Apply the correction?

Yes — prior is outdatedNo — both are true in different contexts

The alternative: the stale belief sits in the graph until someone hits it during an incident. The scorer catches it at the moment the new fact is learned — before it causes a production miss.

How it works

Many roles, many repos, one agent

Each engineer’s Claude Code binds a personal OAuth token to the team’s shared agent identity. One click in the console creates the agent, the scope, and the keys.

PMClaude Desktop
OAuth
BackendClaude Code
OAuth
FrontendClaude Code
OAuth

Team agent

agent_platform_team

Single agentId, shared across everyone

Shared scope

team:platform

memories · beliefs · entities

read + write for every session

1

Create the hive-mind

One form in the Headkey console creates the team agent, the shared scope, and a one-time service-account key — atomically.

2

Each engineer connects

Add Headkey MCP in Claude Code. OAuth opens once, picks the team agent, and stores a revocable bearer token.

3

Shared memory is on

Every learn writes into the team scope. Every ask reads from it. Conflicts surface when the scorer isn't confident.

Stop re-explaining your platform every session.

Create a hive-mind agent in 60 seconds. Connect Claude Code via OAuth. Your team’s next decision builds on its last.