As of 2026-11-17. Source: Maryam Miradi, LinkedIn, 2026-10-23.
Bottom line. A fair 2024 reading list, a weak 2026 stack. Of 19 repos, 12 had a push in the last month, 4 are quiet (5 to 10 months), and 3 have had no push in over a year. Six have moved to a new owner or name. The four categories mix layers, and the list leaves out the frameworks, coding agents, eval and tracing tools most teams actually run.
The list, checked
GitHub stars and last push from the GitHub search API on 2026-11-17. Status is ours.
| Post category | Repo (current home) | Stars | Last push | License | Status |
|---|---|---|---|---|---|
| Multi-agent | camel-ai/camel | 18.0k | 2026-11-16 | Apache-2.0 | Active. Research-flavoured role-play multi-agent framework. |
| Multi-agent | FoundationAgents/MetaGPT (was geekan) | 71.3k | 2026-01-21 | MIT | Quiet ~10 months; moved. SOP-driven "software company" simulation. |
| Multi-agent | Eigenwise/atomic-agents (was BrainBlend-AI) | 6.3k | 2026-11-14 | MIT | Active; moved. Schema-first small components, not an orchestrator. |
| Multi-agent | metauto-ai/GPTSwarm | 1.1k | 2026-02-05 | MIT | Research code (agents as optimisable graphs); quiet ~9 months. |
| Multi-agent | apache/burr (was DAGWorks-Inc) | 2.6k | 2026-11-15 | Apache-2.0 | Active; now Apache. A state machine for one app with a tracking UI. |
| Builders | SWE-agent/SWE-agent | 20.8k | 2026-11-16 | MIT | Active. A research harness for issue fixing and benchmarks. |
| Builders | OpenHands/OpenHands (was All-Hands-AI) | 93.6k | 2026-11-17 | MIT | Active; moved. The strongest pick in the list. |
| Builders | agent0ai/agent-zero (was frdel) | 19.9k | 2026-11-17 | not SPDX | Active; moved. License is custom, so check before embedding. |
| Builders | stitionai/devika | 19.5k | 2025-09-25 | MIT | Stale. README says early, many features broken, points to a successor. Drop. |
| Builders | plandex-ai/plandex | 15.9k | 2025-10-03 | MIT | Stale. Plandex Cloud wound down 2025-10-03; self-host only. |
| Memory & reasoning | stanfordnlp/dspy | 39.1k | 2026-11-13 | MIT | Active, but a prompt and program optimiser, not memory. |
| Memory & reasoning | mem0ai/mem0 | 69.1k | 2026-11-17 | Apache-2.0 | Active. |
| Memory & reasoning | MemGPT → letta-ai/letta → letta-ai/letta-code | 25.5k / 3.9k | 2026-10-18 / 2026-11-17 | Apache-2.0 | Link is two hops stale: letta's README says current source is letta-code; the V1 server is on an archive branch. |
| Memory & reasoning | stanford-oval/storm | 31.9k | 2025-09-30 | MIT | Stale. A research-report writer, not a memory or reasoning component. |
| Tools & observability | Skyvern-AI/skyvern | 23.4k | 2026-11-17 | AGPL-3.0 | Active. AGPL matters if you ship it inside a product. |
| Tools & observability | e2b-dev/E2B | 14.9k | 2026-11-16 | Apache-2.0 | Active. Sandbox SDK; the hosted runtime is paid beyond free tier. |
| Tools & observability | AgentOps-AI/agentops | 6.0k | 2026-06-25 | MIT | Quiet ~5 months. Vendor SDK rather than OpenTelemetry. |
| Tools & observability | ComposioHQ/composio | 31.1k | 2026-11-17 | MIT | Active. Tool integrations and auth; the useful part is hosted. |
GitHub redirects the moved repos, so links still work, but a pinned dependency on the old name can confuse supply-chain tooling.
Corrections
- "Multi-agent orchestration" mixes three things. CAMEL and MetaGPT are multi-agent simulations; Atomic Agents is a component library; Burr is code-owned control flow for one app (closer to durable workflows, R015/R079 territory). GPTSwarm is a paper's code.
- "Memory & reasoning" holds one memory store and a half. mem0 and Letta are memory; DSPy optimises prompts and programs; STORM is an application that writes cited reports.
- "Autonomous builders" are coding agents from the 2024 SWE-bench wave. The coding agents teams run daily in 2026 (Claude Code, Codex CLI, Aider, Cline, OpenCode, Goose) are absent; only OpenHands and SWE-agent are still central.
- "Total cost: $0.00" is not true in use. Model tokens dominate cost; E2B, Composio and Skyvern have paid hosted tiers; AGPL (Skyvern) and a custom license (Agent Zero) carry legal cost.
- Stars are not fitness. Devika (19.5k) and Plandex (15.9k) out-star half the active entries and are stale.
Missing layers
- Agent frameworks most production code uses: LangGraph, OpenAI Agents SDK, Pydantic AI, Google ADK, CrewAI (R081 lists these).
- Durable execution: Temporal, Restate, Inngest, or LangGraph checkpoints (R094).
- Evals: Inspect AI, promptfoo, DeepEval (R099, R107).
- Tracing on OpenTelemetry: Langfuse, Arize Phoenix, OpenLLMetry, instead of a vendor-only SDK.
- Protocols: MCP SDKs and servers, A2A (theme 6.4).
- Sandboxes beyond E2B: Daytona, gVisor/Firecracker-based runners (R093).
- Guardrails and policy: action allowlists and an MCP gateway (R071); interpreter-level sandboxes (R061).
Review checklist for picking from a list like this
- Last push and last release within 90 days; open-issue response time.
- Canonical repo: follow redirects and pin the new owner.
- License fits how you ship (AGPL, custom, "source available").
- Who owns control flow: your code or the model. Prefer code when the steps are known.
- Traces export to OpenTelemetry.
- Sandbox and permission story for any tool that runs code or a browser.
- An eval hook: can you run the same task suite before and after an upgrade?
- Exit path: what breaks if the project stalls (Devika, Plandex).
Caveats
- The post funnels to the author's course; the "20 years, 400+ agents" claims are unverified.
- Stars and dates are as of 2026-11-17.
- Status notes on Devika, Plandex and Letta come from their READMEs; the rest from repo metadata and general knowledge.
Related
R081 (LLM Engineer Toolkit, a larger library list), R083 and R079 (same author: memory blocks, design patterns), R073 (another curated GitHub list), R023 and R095 (build-your-own harness courses), R094 (durable execution), R099 and R107 (evals).
Sources
- LinkedIn post
- GitHub repo pages linked in the table; READMEs of devika, plandex, letta.