About · landing 0.2.2 · techtide-swarm 0.2.2

An organizational ontology for agents, not a headcount fantasy.

Swarm 357 is 357 distinct agent identities across six business layers plus management meta-agents. The Conductor routes work; budgets, bash policy, and portable memory are first-class. We do not claim 357 simultaneous Opus sessions.

Maturity claims below mirror core status documentation against techtide-swarm 0.2.2. Acceptance checks: verification scorecard.

Get startedpip install techtide-swarm==0.2.2

Open source

Swarm 357 ships as open core under Apache-2.0. The runtime and the product surface live in separate public repositories on independent release trains, so memory and orchestration changes never wait on a marketing deploy.

TechTideOhio/swarm-357Core runtime

Python package, Rust Memvid bridge, 357-role roster, soul templates, eval harness.

TechTideOhio/swarm-357-siteProduct surface

This Next.js landing page, the documentation library, and the live demo proxy.

techtide-swarm on PyPIDistribution

Published wheel and source distribution for version 0.2.2.

Tech stack

No hidden platform layer. Everything the runtime depends on is public, documented, and swappable by the operator.

Core runtime

Python 3.10+Runtime for the agent package and CLI
FastAPIHTTP API surface behind swarm serve
Pydantic v2Config and payload validation
SQLiteDurable run checkpoints for inspect and resume
RustMemvid bridge binary
memvid-coreSingle-file .mv2 memory format
Anthropic ClaudeDefault model provider
OpenRouterAlternate provider used for eval baselines

Product surface

Next.js 16App Router, server components, static docs
React 19Component runtime
TypeScriptStrict mode across the codebase
Tailwind CSS v4CSS-first tokens through @theme inline
MDXDocumentation library generated from the core repo
MotionAnimation presets with reduced-motion fallbacks
BunPackage manager, scripts, and build
RailwayHosting for the frontend and the API

Architecture

Management conducts six domain layers. Memory (flat topics plus optional Memvid .mv2) and cost controls cut across every run.

Swarm 357 architecture diagram

Eval evidence

Latest baseline: 154 executions, 145 passed gates, average combined score 0.923, $4.99 of a $5.00 hard cap on openrouter using anthropic/claude-sonnet-4 agents with a Haiku judge. Single-agent 142/142. Swarm 4/12. The wall-clock timeouts behind the swarm number are reported, not hidden.

Eval results by layer: singles 142/142, swarm 4/12

These figures are read from the committed baseline rather than written by hand. Methodology and reproduce steps: eval methodology.

Feature maturity

Agent + Swarm orchestrationBeta
CLIStable / Beta
Memory (flat-file)Stable
Memvid .mv2 via bridgeBeta
BashSecurityGate + argv policyStable
Bash HITL approvalsBeta
SSE event stream (auth + close)Beta
Cost controlsBeta
Eval harnessBeta
HTTP APIBeta
Dream cycleExperimental
Opik cloud observabilityNot implemented

Source of truth: status documentation. Opik cloud observability is not implemented. Local JSONL is the trace store.

Who builds this

Swarm 357 is built by TechTide AI, an automation studio that runs agent workloads in production for its own clients. The features that exist are the ones that survived that use: budget caps, a bash policy gate, durable checkpoints, and memory that outlives a process.

  • Licensed Apache-2.0. Contributions are welcome through the contributing guide.
  • The interface follows a documented design system, enforced in continuous integration.
  • Package tests and a coverage floor run on every change to techtide-swarm 0.2.2.
  • BashSecurityGate ships 13 pattern rules covered by 33 scenario cases.
  • Release history is published in the changelog.