Reference

Roadmap — scale to 100+ agents

Honest status first: this is decided-as-roadmap, not implemented runtimes. It doesn't ship daemons. Everything below is parked as a future Seed, to be unlocked only when a product actually needs it.

Line drawing of a scattered swarm of bees

Target operating model

Many products, times many build-master instances, times many pool instances, times many concurrent units — equals one protocol. Every agent still cold-starts from Orchestration. Density increases as a program grows; the process itself never forks into something new.

Swarm capacity and leases

Concern Rule
Concurrent buildsOne worktree / unit / PR; many units OK across branches
CLOSEOUTOne atomic build-master lease for shared finalization
PoolsBrains write insight only; Admin promotes
Cross-pool@mention + INBOX routing — no race CP-0
Cross-hiveLocal project only; file/agent share needs a human Admin gate
Quality under loadNever skip audit, preview, or program lock to "go faster"

Throughput is concurrent units that still pass every gate — not agent count alone.

Brain scheduling conventions

Mode When
Human-pacedDefault — Admin asks, or a CLOSEOUT digest triggers a skim
Light recurringOptional Admin cadence per pool — still Cursor search, no silent OpenRouter
Automation-eligible CP-0Only where the product's MASTER-PLAN CP-0 mode allows it and CP0-POLICY checks PASS

Quality under load

  • The same cold-start serves a one-agent docs unit and a hundred-agent program.
  • Role owns / must-not lines never dilute as a swarm grows.
  • A bee roster is required on every new build system, no exceptions.
  • The local project boundary holds — cross-hive share only with a human Admin gate.
  • Proposal → ship conversion always requires an Admin promote.
  • Doctrine drift is caught via CLOSEOUT digests flowing into hive/proposals/.
  • Cross-pool collisions escalate to INBOX; Admin decides.

Success metrics

Metric Intent
Time-to-cold-startA cold agent finds its role and next action from hive + instance alone
CLOSEOUT completenessDigests, history, and dual worklogs present — or honestly deferred
Proposal → ship conversionAdmin-promoted cards that become landed units
Doctrine drift rateUnplanned hive edits vs. accepted proposals
Cross-pool collision rateCompeting promotes or unresolved INBOX races

Parked tooling seeds — not built here

Each becomes its own Seed and blueprint only when Admin unlocks it:

Parked seed Job
Digest aggregatorRoll up pool DIGESTS across a product into one skim
Pattern-upgrade docs botDraft hive/proposals/ from digest residuals — docs-only, Steward accepts
Lease dashboardShow active CLOSEOUT leases / worktrees — observability only
Optional CP-0 runnerOnly for automation-eligible instances — still obeys CP0-POLICY
Shared lifecycle-UI packageBloom meaning #3 UI, kept separate from build ops
Measure / observabilityInstrumentation for the Demand → … → Evolve Measure stage
Admin-agent Steward lockParked future option — Admin-agent powers only while a human Steward lock is unlocked

Canonical source: HIVE-ROADMAP.md — see also the public hive map and the parked Admin-agent lock note it links to.