Pillar · Agentic Swarms

Hundreds of Agents. One Machine. Zero Cloud.

Other swarm tools spin your agents up in a configured cloud environment. Anvaya's swarm layer runs on your own hardware: 25 agents ran in parallel on a Core 2 Duo with 4 GB of RAM, no crashes, no faults.

Direct answer

An agentic swarm is a supervised group of coding agents working one job: a planner partitions it, agents execute in isolation, and verification decides what lands. Anvaya Swarm is built around one idea, that agents should be cheap enough to run by the hundred on hardware you already own. Each agent is its own process under a supervisor: 24 concurrent by default with a 256-agent hard cap, placement in shared, git-worktree or cloned workspaces, and builtin bundles for general swarms, defect fan-outs, MVP builds and refactors. It is light enough to multiply: a 32-agent soak cost 103 MB total agent RSS, and 25 sessions ran in parallel on a Core 2 Duo with 4 GB of RAM. Completion is not a model self-report: Tier 0 checks (linters, tests, LSP, dry runs) run first, failures escalate to a Tier 1 critic, and approved work lands only on anv/integration/<run>, never pushed, under any policy. A production E2E delivered three SaaS apps with zero merge conflicts.

Anatomy

What The Supervisor Actually Does.

Eight shipped mechanisms, live-verified at N=32.

01Supervisor, not a monolithThe supervisor is a separate process; each agent is its own process. The CLI attaches with anv --swarm <sock>; administration is anv swarm up/status/logs/board/approve/reject/cancel. Process-per-agent means one runaway agent cannot take the fleet down.

02Partition, then consolidateA partitioner prompt splits a job into tasks with mandatory verification commands; a consolidator reconciles results. The control graph is a guarded state machine with typed artifacts (produces/consumes) rather than a chat fan-out.

03Choose the isolation levelPlacement per run: shared (one workspace, cheapest), worktree (isolated git worktrees), or clone (fully separate checkout). The defect, mvp and refine bundles default to worktree placement.

04Verdict before trustTier 0 verification is zero-token: linters, test runners, the LSP, dry runs. Only failures escalate to a Tier 1 critic. A task is done when an external check says so, never when the agent says so.

05Repair with an escalation ladderFailed tasks retry through a novelty guard and a ladder: same model → repairer role → widen scope → replan → human gate. Repair attempts are capped per task and defaults retry only transient provider errors and crashes.

06Integration is a gate, not a mergeApproved work lands on anv/integration/<run>. Nothing is pushed under any policy, including --yolo. Conflicts and failed verdicts are parked and reported, not silently dropped.

07The safety envelopePer-wave admission checks disk, file descriptors and builds; the supervisor reduces concurrency on 429s; a dead-man switch keyed to verdict-state changes halts a wedged run. Liveness is tracked with a 10-second heartbeat, suspect at 30 s, lost at 60 s, STALLED in the TUI.

08Bounded by default24 agents run concurrently by default with a 256-agent hard cap and a 20/s spawn rate. The telemetry sampler keeps a fixed 512-sample ring per agent and stretches to a 15 s interval when all agents are idle.

Fan-out Or Not

One Agent, Parallel Agents, A Supervised Swarm.

The difference is not the number of agents. It is who holds the ledger, the budget and the merge gate, and whether the fleet fits on your laptop.

DimensionOne agentAd-hoc parallel agentsSupervised swarm

Work splitAgent decides inlineHuman splits by handPartitioner with typed artifacts

IsolationNoneWherever you launched themshared / worktree / clone per run

LivenessOne processEach terminal for itself10 s heartbeat, suspect/lost states

VerificationAgent self-reportManual reviewTier 0 objective checks, then critic

Failure handlingYou retryYou retryNovelty-guarded repair ladder

Landing workYour working treeYour working treeIntegration branch, never pushed

Measured

What We Have Actually Run.

01N=32 supervisor soakDummy agents, 15 s: supervisor RSS max 7,553,024 B, spawn wall 396 ms, status p50 4.2 ms / p99 14.9 ms, pass=true. Registered agents: 32.

02Production E2E, 2026-09-13Three SaaS apps delivered on the mvp bundle with worktree placement; integration worktrees green with 35/43/38 tests; conflicts [], parked [], verdict failures 0.

03Autonomy baselinen-of-3 across three modes, nine runs: all modes 100% first-pass on the corpus; corpus expanded to six entries.

04Not yet measured256-agent soak (deferred by product decision), offline swarm throughput on a single GPU, and cross-vendor swarm comparisons. We do not claim them.

Questions

Asked About Swarms.

What is an agentic swarm?

A group of coding agents working one job under a supervisor: one planner partitions the work, many agents execute it in isolation, and verification plus consolidation decide what lands. It differs from a single agent with subagents in scale and durability, processes, heartbeats, admission control and an integration gate instead of one context window.

How many agents can Anvaya run?

24 concurrent by default, with a 256-agent hard cap and a 20-per-second spawn rate. A 32-agent dummy-agent soak passed with the supervisor at 7.55 MB max RSS and 396 ms spawn wall; a 256-agent soak is deferred by product decision and is not claimed.

Do swarms push code automatically?

No. Integration lands only on anv/integration/<run>; pushing is never performed under any policy including --yolo. You review the integration branch, then merge it yourself.

What stops a confused agent from wrecking the repo?

Placement isolation (worktree or clone), tool-layer sandbox policy, per-task budgets and round limits, the repair ladder capped by novelty guard, and the integration gate. A hung agent triggers graceful cancel → SIGTERM → SIGKILL on the process group.

Is the swarm measured?

Partly, and we label it. Measured: the N=32 supervisor soak, and a production E2E on 2026-09-13 where three SaaS apps were delivered on the mvp bundle with worktree placement, green integration suites (35/43/38 tests) and zero conflicts. Not yet measured: 256-agent scale, offline throughput, and vendor comparisons.

When is a swarm the wrong choice?

Small, tightly coupled edits, two agents editing the same file usually lose to one agent plus a checkpoint. Swarms pay off on partitionable work: many independent modules, audits, migrations, and defect fan-outs where verification is objective.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.