Todas as referências, papers, repos e links que informaram o design de Wake. Organizado por tema.
- Managed Agents Overview — O que é, primitivas, lifecycle, billing, beta access. Beta header
managed-agents-2026-04-01. Rate limits 300/600 rpm. - Agent Setup — Schema completo do Agent: name, model, system, tools, mcp_servers, skills, multiagent, description, metadata. Versionamento, archive.
- Environment Setup — Container template, package managers (apt/cargo/gem/go/npm/pip), networking (unrestricted/limited), allowed_hosts, allow_package_managers, allow_mcp_servers.
- Sessions — Two-step lifecycle (create → send event), pin to version, vault_ids, statuses (idle/running/rescheduling/terminated).
- Scaling Managed Agents: Decoupling the brain from the hands — O paper-chave. Descreve brain/hands split, session log durável, harness stateless com wake/getEvents/emitEvent, sandbox como tool, vault + proxy. TTFT p50 -60%, p95 -90%.
- Building agents with the Claude Agent SDK — Filosofia "give your agents a computer." Agent loop: gather context → take action → verify work → repeat. Verificação: rules-based, visual, LLM-as-judge.
- Making Claude Code more secure and autonomous with sandboxing — Threat: prompt injection. Filesystem isolation + network isolation. Reduz prompts de permissão em 84% internamente.
- Claude Code auto mode: a safer way to skip permissions — Two-layer defense: prompt-injection probe + transcript classifier (Sonnet 4.6). Stage 1: 8.5% FP / 6.6% FN. Full pipeline: 0.4% FP / 17% FN.
- Tool use with Claude — Client vs server tools, tool_choice options, agentic loop, fine-grained streaming.
strict: truepara conformância de schema. - Code execution tool —
code_execution_20250825ecode_execution_20260120. Container 5GB/5GB/1cpu, internet disabled, 30-day expiration, 1550 hours free/month then $0.05/h. Multi-environment warning explícito. - Model Context Protocol — "USB-C for AI applications." Open protocol para conectar AI a data sources, tools, workflows.
- anthropic-experimental/sandbox-runtime — Componente reusável central. OS-level filesystem + network sandboxing. macOS via sandbox-exec/seatbelt, Linux via bubblewrap + seccomp BPF + Unix socket relay via socat. HTTP/HTTPS proxy + SOCKS5 proxy.
- Infisical Agent Vault — MITM HTTPS proxy que substitui placeholders por credenciais reais. Integra com Claude Code, OpenClaw, Hermes. Reusable componente para o vault layer de Wake.
- OpenClaw Managed Agents — Clone explícito Managed Agents OSS. MIT. ~410 stars. v0.2.0 Apr 2026.
- OpenHands V1 SDK — Coding agent com EventLog imutável. Paper: arxiv 2511.03690. 68.6k stars. MIT. $18.8M Series A.
- Multica — Control plane sobre CLIs (Claude Code, Codex, Gemini CLI). 10.7k stars. Q1-Q2 2026 momentum.
- Microsoft Agent Framework — v1.0 GA Apr 2026. MIT. Python + .NET. Checkpointing, MCP, A2A, multi-provider.
- LangChain Deep Agents Deploy — Explicit "open alternative to Claude Managed Agents." Self-host via LangSmith Deployments.
- Kitaru (ZenML) — Python decorator-based durable execution.
@flow/@checkpoint,kitaru.wait()pause/resume, artifact lineage. - Inside Claude Managed Agents (Pluto Security writeup) — Análise externa da arquitetura Managed Agents.
- LangGraph — Graph-based agent framework, durable execution via checkpointers.
- CrewAI — Multi-agent orchestration, 30k+ stars, MIT.
- Pydantic AI — Type-safe agents, native MCP/A2A/durable-execution.
- Claude Agent SDK — Anthropic SDK.
- LangChain — 100k+ stars. Geral framework.
- LlamaIndex — Data + agent framework.
- Microsoft AutoGen → Agent Framework — Production successor.
- AG2 (ex-AutoGen) — Multi-framework interop.
- OpenAI Agents SDK — OpenAI's SDK.
- Strands Agents — AWS, model-driven, MCP-native.
- Dapr Agents — Apache 2.0, Dapr-backed.
- Pydantic AI — Type-safe.
- SmolAgents (HuggingFace) — Small, hackable.
- Agno (ex-Phidata) — "AgentOS" production.
- Mastra — TypeScript.
- BeeAI (IBM, LF) — Linux Foundation.
- OpenHands (ex-OpenDevin) — 68.6k stars.
- goose (Block) — Rust, 32k stars, MCP-first.
- Aider — Terminal, git-first.
- Cline — VS Code extension.
- Continue.dev — IDE-integrated.
- Tabby — Self-hosted code assistant.
- Open Interpreter — Local code execution.
- SWE-agent — Eval-focused.
- Temporal — Workflow engine genérico, multi-language. Integração com OpenAI Agents SDK GA Mar 2026.
- Restate — Durable handlers, journal-backed step replay, Pydantic AI integration.
- DBOS — Postgres/SQLite-backed durable workflows.
- Inngest — Event-driven durable execution.
- Inngest AgentKit — TypeScript multi-agent networks.
- Trigger.dev — TS-first durable workflows.
- Conductor OSS — Netflix event-driven, Apache 2.0.
- Cloudflare Workflows — Platform-specific.
- E2B — Firecracker microVMs, ~150ms cold start, open core MIT.
- Daytona — Docker-based, sub-90ms cold start.
- Modal — Serverless GPU + sandbox.
- Vercel Sandbox — Firecracker microVMs.
- Riza — Code interpreter alternative.
- Pyodide — Python in WASM.
- Letta (ex-MemGPT) — Memory-first agent runtime, three-tier core/recall/archival.
- Mem0 — Universal memory layer, 21+ framework integrations, Apache 2.0.
- Zep / Graphiti — Temporal knowledge graph.
- Cognee — GraphRAG, deep knowledge retrieval.
- Langfuse — OSS leader, ClickHouse Series D $400M Jan 2026.
- Arize Phoenix — Elastic 2.0, OpenInference-native.
- Helicone — Proxy-based, simplest install.
- Braintrust — Eval-first.
- W&B Weave — ML obs platform.
- Literal AI — LLM obs.
- Lunary — Open core.
- Laminar — Agent rollout debugger.
- LangSmith — Closed, LangChain-native.
- LiteLLM — OpenAI-compatible across 100+ providers. Componente reusable central.
- OpenRouter — SaaS marketplace.
- Portkey — Observability gateway.
- Vercel AI Gateway — Provider abstraction.
- agentgateway (Linux Foundation) — AI-native gateway, Rust, MCP + A2A + LLM gateway. Componente reusable.
- Model Context Protocol (MCP) — Anthropic + comunidade. Standard para tools.
- Agent-to-Agent Protocol (A2A) — Google. Standard para multi-agent.
- Open Agent Specification — Oracle + Microsoft + Google. Declarative agent definition format.
- OpenTelemetry — CNCF. Tracing standard.
- OpenInference — OpenTelemetry semantic conventions for LLM/agent.
- OpenHands V1 SDK paper — Architecture blueprint quase idêntico à tese Wake.
- Voyager: An Open-Ended Embodied Agent with LLMs — Skill library evolution.
- ReAct: Synergizing Reasoning and Acting in LLMs — Tool use loop padrão.
- Reflexion: Language Agents with Verbal Reinforcement Learning — Self-correction loops.
- SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering — Coding agent design.
- "Designing Data-Intensive Applications" (Kleppmann) — Capítulos 7, 11.
- "Event Sourcing" (Fowler) — Fundamentos.
- LangChain blog: Deep Agents Deploy — Posicionamento explícito como alternativa OSS.
- Anthropic Cookbook — Patterns oficiais.
- MCP Server Registry — Lista de MCP servers.
- Anthropic lança Claude Code (2024)
- MCP é open-sourced (Nov 2024)
- OpenHands ganha Series A $18.8M (2025)
- Anthropic publica engineering post sobre Managed Agents (2026)
- Anthropic open-sources sandbox-runtime (2026)
- Multica explode com 10k+ stars Q1-Q2 2026
- Microsoft Agent Framework v1.0 GA (Apr 2026)
- LangChain Deep Agents Deploy lançado (Q1 2026)
- Langfuse acquired by ClickHouse $400M Series D (Jan 2026)
- Open Agent Specification publicado por Oracle + MS + Google (2026)
Atualizada conforme novos projetos / papers / posts aparecerem. PRs adicionando recursos relevantes são bem-vindos.
Critério para inclusão: o recurso deve ter impactado ou influenciado decisão de design em Wake, ou ser potencial componente reusável, ou ser competidor/adjacente a tracking.