The Agentic Engineer
- OpenCode hit 120K GitHub stars and 5M monthly devs, making it the largest open-source coding agent ever shipped. It works with 75+ LLM providers and your existing Copilot or ChatGPT Plus subscription.
- Claude Code now accepts push events mid-session via Channels. CI results, webhooks, and chat messages flow into running agents. Coding agents just went from request-response to event-driven.
- Hindsight, a new agent memory system, hit SOTA on LongMemEval by making agents learn from experience, not just recall conversations. Two lines of code to integrate. Docker one-liner to deploy.
OpenCode: 120K Stars and the End of the Closed Coding Agent
The numbers are hard to argue with. OpenCode crossed 120,000 GitHub stars this week, with 5 million monthly active developers and 800 contributors pushing 10,000+ commits. The Hacker News thread pulled 1,237 points and 614 comments. For context, that puts it ahead of VS Code's first-year trajectory.
OpenCode is an open-source coding agent that runs in your terminal, on your desktop, or inside your IDE. It supports 75+ LLM providers out of the box. You can log in with your existing GitHub Copilot or ChatGPT Plus subscription. No new API key required.
The technical story is what makes it interesting. OpenCode ships with full LSP (Language Server Protocol) support, which means it understands your codebase the way your editor does. Jump to definition, find references, rename symbols. Multi-session support lets you run parallel agent tasks across different parts of a project without them stepping on each other.
Look, the coding agent space has been dominated by closed-source tools. Cursor, Claude Code, GitHub Copilot Workspace. They're good. But they're black boxes with monthly subscriptions and usage caps. OpenCode's pitch is simple: bring your own model, own your workflow, inspect every line of the agent's reasoning.
The 614-comment HN thread is worth reading. The dominant take: open-source coding agents don't need to be better than Claude Code. They need to be good enough, transparent, and free from vendor lock-in. At 120K stars, the community has voted.
There's a practical angle here too. If you're building internal developer tools or need a coding agent that runs on-prem (compliance, air-gapped environments, regulated industries), OpenCode is the first credible option at this scale. The LSP integration means it slots into existing toolchains without ripping anything out.
Open-source coding agents already matter. The real pressure is on closed-source incumbents to justify their pricing when the open alternative has 800 people improving it for free. My bet: they'll have to compete on UX and speed, because the capability gap is closing fast.
Source: opencode.ai | HN Discussion (1,237 pts)
Claude Code Channels: Push Events Into Running Agent Sessions
Anthropic shipped Channels for Claude Code. MCP servers can now push events (CI results, chat messages, webhooks) into a running session. Telegram and Discord bridges included. Two-way: Claude reads events and replies through the same channel. This turns coding agents from "ask a question, get an answer" into event-driven systems that react to their environment. The always-on agent pattern just went mainstream.
Source: Hacker News (398 pts)
Flash-MoE: 397B Parameters on a MacBook at 4.4 tok/s
Flash-MoE streams Qwen3.5-397B-A17B (209GB) from SSD through a custom Metal compute pipeline on a MacBook Pro with 48GB RAM. Pure C, hand-tuned Metal shaders, FMA-optimized dequant kernels. Full tool calling at 4-bit quantization. No Python, no frameworks. The key insight: trust the OS page cache instead of trying to fit everything in memory. Built in 24 hours by a human-AI pair. Local agent inference with a 397B MoE model is no longer theoretical.
Source: Hacker News (353 pts)
NVIDIA NemoClaw: Sandboxed Runtime for Always-On Agents
NVIDIA open-sourced NemoClaw, a reference stack for running OpenClaw agents inside Landlock + seccomp + network namespace sandboxes with managed Nemotron inference. Part of the NVIDIA Agent Toolkit. Alpha release, DGX Spark support included. The sandbox approach (kernel-level isolation, not just Docker) is the right answer to "should I let an AI agent run shell commands on my production machine?"
Source: Hacker News (383 pts)
Mistral Forge: Train Frontier Models on Your Data
Mistral Forge launched enterprise custom model training: pre-training, post-training, and RL on proprietary data. Partners include ASML, ESA, Ericsson, and DSO Singapore. The agent angle matters most. RL-tuned models make tool selection more precise and multi-step workflows more reliable. If your agents keep picking the wrong tool on step 3 of a 5-step chain, this is the fix. 730 HN points suggest the market agrees.
Source: Hacker News (730 pts)
OpenViking: ByteDance's Context Database for AI Agents
ByteDance/Volcengine open-sourced OpenViking, a context database that uses a filesystem model to unify agent memory, resources, and skills. L0/L1/L2 tiered context loading cuts token consumption. Directory recursive retrieval beats flat RAG in their benchmarks. +6,297 stars in one week (18,203 total). The visualized retrieval trajectories for debugging are a nice touch for anyone tired of guessing why their agent hallucinated.
Source: GitHub Trending (+6,297 stars/week)
Attention Residuals: A Drop-In Transformer Upgrade
Paper: MoonshotAI/Attention-Residuals (arXiv + code)
The problem. Standard transformers use residual connections that just add each layer's output to the running total. PreNorm (the variant everyone uses) has a known flaw: the residual stream's magnitude grows without bound as you stack layers. Deeper models get harder to train. This has been an open problem since 2020.
The fix. Replace the fixed residual addition with learned, input-dependent attention over depth. Each layer doesn't just add its output. It looks back at all previous layers and decides how much of each to keep. Think of it as giving each layer a vote on which earlier representations still matter.
Why it works. The attention mechanism (softmax over depth) naturally bounds the magnitude. No more unbounded growth. The model learns which layers to trust for different inputs. A math token might lean heavily on early syntactic layers. A reasoning token might weight the deeper abstract layers.
The practical version. Full Attention Residuals adds overhead. But Block AttnRes groups layers into blocks and only applies depth-attention at block boundaries. Marginal compute cost, measurable quality gains. Drop-in replacement for standard residual connections. No architecture changes needed.
Why builders should care. Every foundation model uses residual connections. If Block AttnRes becomes standard (and the results suggest it should), the next generation of base models will be more trainable at depth. That means better reasoning, longer chains of thought, and more reliable multi-step agent behavior. You won't integrate this yourself, but you'll benefit when model providers do.
Time saved: 5 min read vs 35 min paper. 7.0x compression.
Hindsight: Agent Memory That Actually Learns
Hindsight is a memory system for AI agents that hit #1 on the LongMemEval benchmark, independently verified by Virginia Tech. It's not another RAG wrapper. The difference: a retain/recall/reflect loop that makes agents learn from experience, not just store conversations.
Most agent memory systems are glorified search indexes. You stuff conversation history into a vector store and hope cosine similarity finds the right context. Hindsight adds a reflection step. After each interaction, it extracts lessons, updates its model of the user, and consolidates related memories. Over time, the agent gets better at predicting what you need.
Integration is minimal:
from hindsight import Hindsight memory = Hindsight() # Wrap your LLM calls response = memory.chat(model="gpt-4", messages=messages)
Deploy with Docker:
docker run -d -p 8100:8100 vectorize/hindsight:latest
The benchmark numbers matter. On LongMemEval (the standard for long-term agent memory), Hindsight beats both vanilla RAG and knowledge graph approaches. Fortune 500 companies are running it in production. The API is simple enough that you can swap it into an existing agent in under an hour.
If your agents forget what happened two conversations ago, this is worth 30 minutes of your time.
Stars: 5,788 (+1,771/week) | License: Apache 2.0
Framework Star Tracker
Weekly star tracker, March 25, 2026. Deltas vs. Issue #4 (March 18).
| Framework | Stars | Weekly Δ |
|---|---|---|
| OpenClaw | 331,202 | +14,913 |
| n8n | 180,616 | +1,248 |
| Dify | 134,078 | +1,073 |
| LangChain | 130,704 | +1,000 |
| AutoGen | 56,058 | +382 |
| Flowise | 51,001 | +211 |
| LlamaIndex | 47,893 | +194 |
| CrewAI | 46,937 | +731 |
| Semantic Kernel | 27,530 | +62 |
| LangGraph | 27,215 | +698 |
| Haystack | 24,590 | +70 |
| Vercel AI SDK | 22,926 | +259 |
| Mastra | 22,241 | +206 |
| OpenAI Agents SDK | 20,212 | +180 |
| Strands Agents | 5,355 | +44 |
Notable moves: OpenClaw added another 14.9K stars this week (331K total), pulling further ahead of n8n (180K). LangChain crossed the +1K weekly delta for the first time in months. CrewAI (+731) and LangGraph (+698) are the fastest movers in the mid-tier. Mastra (22.2K) continues closing the gap on Vercel AI SDK (22.9K). The top 4 now account for more combined stars than the bottom 11. Consolidation at the top, fragmentation everywhere else.
NousResearch shipped Hermes Agent this week with 10,698 stars and a literal hermes claw migrate command to poach OpenClaw users. Self-improving skills, multi-platform gateway, any model via OpenRouter. Meanwhile Claude Code Channels just turned coding agents into event-driven systems. The always-on agent space is about to fragment the way web frameworks did in 2015. Everyone will pick a side, nobody will agree, and in 18 months we'll pretend we always knew which one would win. My money is on the one that ships the best memory system, not the best model integration. Memory is the moat. Everything else is plumbing.
What's your primary coding agent right now?
| 👻 Kiro |
| 🤖 Claude Code |
| ⚡ OpenCode |
| 🐙 GitHub Copilot |
| 🔧 Other / None |
CertPrep
17,000+ practice questions across 49 certification exams — Azure, GCP, CISSP, CompTIA, Cisco, Kubernetes & more. Timed tests, detailed explanations for every option, progress tracking. Free tier for every exam. One-time purchase, no subscriptions.
Download FreeWant to sponsor this newsletter? Get in touch
Like what you read?
Forward this to a friend who's building with agents.
Subscribe to The Agentic Engineer