Open-source layered multi-agent orchestration

357 agent identities. 6 business layers.
One bounded execution model.

Python · Anthropic Claude · portable .mv2 memory via a Rust bridge · per-layer cost controls · BashSecurityGate.

Read the source →$pip install -e 'packages/techtide-swarm[dev]'

357: agent identities
7: layers (1 meta + 6 domain)
13: security patterns

The honest bit

What "357" actually means

357 agents means 357 distinct YAML roster entries, materialized at runtime — for example sales-outreach-specialist-015. It is not a claim about 357 parallel long-running LLM sessions. Layer execution uses bounded concurrency.

The roster is generated by a script and verified in docs/VERIFY.md. Inspect both before trusting the count.

docs/VERIFY.md Roster generator script

Architecture

A meta-layer that conducts six business layers.

The Conductor reads strategy and curated memory, then dispatches tasks downward. Each business layer holds its own roster, budget, and memory store.

10 management agents · 347 domain agents · 357 total identities

What's different

Four design decisions you can audit.

Portable .mv2 memory

Single-file Memvid store per agent layer. WAL-based crash safety, full-text and vector search. No database server to run.

Business-layer ontology

Six domain layers plus management meta-agents. Not just 'agent A talks to agent B' — work has a place to live.

Layered cost controls

Per-agent budget caps. Per-layer daily limits. Automatic model downgrade at 80% utilization.

BashSecurityGate

13-pattern regex validator wired into the Bash tool. Blocks destructive commands, secret exfiltration, dangerous system ops. 50+ tests.

Command line

One CLI. Eight verbs.

The whole framework is reachable from swarm. Dashboards, planning, evaluation and the memory dream cycle live in the same tool.

~/swarm

$ swarm run <task>                        # routes through Conductor
$ swarm boot                                    # loads 357-agent roster
$ swarm agent --list                    # all agents
$ swarm dream                                  # memory consolidation
$ swarm plan <task>                      # deep planning via Claude Opus
$ swarm eval                                    # 5-task benchmark
$ swarm status                                # runtime dashboard
$ swarm cost                                    # spend dashboard

Feature maturity

Where each piece actually is.

No "production-ready" gloss. Build on the Stable bits, experiment with Beta, expect changes on Alpha, treat Stub as a placeholder.

Feature	Status	Notes
Agent runtime	Beta	Core abstraction stabilizing
Swarm orchestrator	Beta	Conductor + bounded layer execution
CLI	Stable	Most verbs stable; a few subcommands Beta
Memory (.mv2)	Beta	WAL crash-safety landed; APIs may shift
BashSecurityGate	Stable	13 patterns, 50+ tests
Cost controls	Beta	Per-agent + per-layer caps; downgrades wired
Dream cycle	Alpha	Memory consolidation, expect rough edges
Eval harness	Alpha	5-task benchmark, scoring evolving
WebSearch tool	Stub	Interface in place, provider not wired

Honest comparison

vs. LangGraph, CrewAI, OpenAI Agents SDK.

The other frameworks are good tools. Swarm 357 picks a narrower opinion: layered business ontology, portable single-file memory, per-layer cost controls.

	Swarm 357	LangGraph	CrewAI	OpenAI Agents SDK
Business-layer ontology	Six domain layers + meta	Graph of nodes	Crew/role free-form	Agent + tools
Memory store	Portable .mv2, single file	Bring your own	Bring your own	Threads / vector store
Per-layer cost caps	Built in, auto-downgrade	DIY	DIY	Per-run limits
Bounded layer execution	First-class	Manual via graph	Manual	Manual
Bash security gate	Built in, 13 patterns	Not included	Not included	Not included
License	Apache 2.0	MIT	MIT	Proprietary SDK

Read it. Run it. Argue with it.

Apache 2.0. Inspect the roster, run the eval, file an issue.

github.com/TechTideOhio/swarm-357 →$pip install -e 'packages/techtide-swarm[dev]'

357 agent identities. 6 business layers.One bounded execution model.