AgentBreeder v2.0: Frameworks, Clouds, Languages, Providers Are Now Plug-Ins

v1 of AgentBreeder shipped a working agent platform. v2 turns it into a substrate — anything above the substrate (frameworks, clouds, languages, providers) plugs in without engine changes.

Six tracks landed at once. Here's what each one buys you.

Track F — Provider Catalog

Adding Nvidia NIM used to mean writing a provider class. Same for Kimi K2, Groq, Together, Fireworks. They're all OpenAI-compatible — only base_url and api_key differ. v2 collapses them into one generic class plus a checked-in catalog YAML with 9 presets out of the box.

model:
  primary: nvidia/meta-llama-3.1-405b-instruct

Not on the list? agentbreeder provider add my-vllm --type openai_compatible --base-url http://… --api-key-env MY_KEY and you're done.

Track G — Model Lifecycle

Provider model lists change weekly. v2 auto-discovers from each provider's /models endpoint, marks new ones active, deprecated ones deprecated, and after 30 days of absence: retired. Audit events for every transition. CLI: agentbreeder model sync.

Track H — Gateways as First-Class

LiteLLM and OpenRouter graduate from "connector tucked away in a folder" to first-class catalog entries with type: gateway. Three-segment refs route the request:

model:
  primary: openrouter/moonshotai/kimi-k2

Switch upstream providers without touching agent code.

Track I — Polyglot Runtime Contract + Go SDK

Every agent container — Python, TypeScript, Go, Kotlin, Rust, .NET — satisfies the same versioned HTTP contract: /health, /invoke, /stream, /resume, /openapi.json, bearer auth via AGENT_AUTH_TOKEN. Spec lives at engine/schema/runtime-contract-v1.md + runtime-contract-v1.openapi.yaml.

The first Tier-2 SDK ships with v2: Go (sdk/go/agentbreeder/). Kotlin, Rust, and .NET are tracked in their own issues for v2.x.

agentbreeder init --lang go --framework custom my-go-agent

Track J — The Sidecar

CLAUDE.md described "the sidecar pattern (planned)" since v1. v2 ships it. A single Go binary auto-injected next to every agent that declares guardrails:, MCP tools:, or a2a:. Handles bearer auth, OTel tracing, cost attribution, PII guardrails, A2A JSON-RPC, MCP passthrough — once, in Go, and every language SDK gets it for free.

Track K — Workspace Secrets

engine/secrets/ had four backends (env, AWS, GCP, Vault) but no workspace binding. v2 adds:

• keychain backend (cross-platform via keyring)
• One backend per workspace (AGENTBREEDER_INSTALL_MODE=team defaults to env, etc.)
• Auto-mirror at deploy — declared secrets: get written to the target cloud's native store under agentbreeder/<agent>/<secret>, runtime SA gets secretAccessor, no plaintext in the image

agentbreeder secret set OPENAI_API_KEY        # one command, right backend per workspace

What didn't ship (yet)

We were honest about scope. v2 doesn't include:

• Track A — Workspace primitive (#146): workspace.yaml, agentbreeder workspace init/login, multi-cloud config block. The pieces that depend on it (gateway config, secrets backend) ship in v2 with a per-agent fallback.
• /agent-build redesign (#147): the next-gen entrypoint that walks you from pip install through deploy.
• Polyglot phase 2/3 SDKs: Kotlin (#188), Rust (#189), .NET (#190). Tracked, not yet built.
• Auto-mirror to Azure — the secrets auto-mirror flow currently covers AWS ECS + GCP Cloud Run. Azure Container Apps support is in the next minor.
• Daily model-sync cron in cloud workspaces — Track G ships the engine + CLI; the cron itself is a system-cron sample for v2.0, native scheduler in v2.1.

Try it

pip install agentbreeder==2.0.0
agentbreeder provider list                  # see the 9-preset catalog
agentbreeder secret set OPENAI_API_KEY      # workspace backend, masked input
agentbreeder model sync                     # auto-discover from your configured providers

Or docker compose -f deploy/docker-compose.yml up -d and open http://localhost:3001.

Full changelog: v2.0.0 release notes. Architecture spec: docs/architecture/platform-v2.md.