The Agentic Engineer
- JetBrains Central launches an open control plane for coding agents. 90% of devs use AI at work, but only 13% use it across the full SDLC. Central is their bet on closing that gap.
- ARC-AGI-3 drops and frontier AI scores below 1%. Humans score 100%. The benchmark that keeps humbling the field just got harder.
- AWS Labs ships Agent Plugins for Claude Code and Cursor. Reusable, versioned skills that let coding agents architect, deploy, and operate on AWS. Six plugins covering serverless, Amplify, databases, and GCP-to-AWS migration.
JetBrains Central: The Control Plane for Agent-Driven Software
JetBrains just made their move. Central is an open orchestration platform that connects any coding agent to any IDE. Claude Code, Codex, Gemini CLI, your custom agent. All of them, one control plane.
The numbers from their developer survey tell the story. 90% of devs use AI at work. 22% already use coding agents. 66% of companies plan to adopt agents within 12 months. But only 13% use AI across the full software development lifecycle.
That 13% number is the one that matters. It means 87% of teams are using AI for code generation and nothing else. No CI/CD integration. No test generation. No deployment automation. Agents are stuck in the editor.
Central's pitch: governance, cost attribution, and shared semantic context across repositories. Think of it as a management layer that sits between your agents and your codebase. Every agent action gets logged. Every token gets attributed to a team or project. Every agent shares the same understanding of your architecture.
This is the "who controls the agents" problem. Right now, most teams have individual developers running Claude Code or Copilot with zero visibility into what those agents are doing across the org. No audit trail. No cost tracking. No way to enforce coding standards at the agent level.
JetBrains is betting that as agent adoption scales from 22% to 66%+, companies will need this layer. And they're probably right. The pattern is familiar: containers needed Kubernetes, microservices needed service meshes, and coding agents need a control plane.
The open architecture is smart. By supporting any agent and any IDE, they avoid the lock-in problem that would kill adoption. A team running VS Code with Claude Code and IntelliJ with Gemini can manage both from one dashboard.
My take: the real competition here is GitHub. Copilot already has enterprise features, but they only work with Copilot. Central works with everything. If JetBrains executes well, they become the Switzerland of coding agents. Neutral ground where every agent vendor plays nice.
The 13%-to-100% gap is where the money is. Central is JetBrains' bet that they can own the infrastructure that closes it.
ARC-AGI-3: Frontier AI Scores Below 1%, Humans Score 100%
Chollet dropped ARC-AGI-3 and it's brutal. Interactive, turn-based environments where agents must explore, infer goals, and plan action sequences without instructions. Frontier models score below 1%. Humans solve every single one. The $1M+ Kaggle prize is still up for grabs. Two years of scaling and prompting tricks, and the gap between human reasoning and AI pattern matching hasn't budged on this benchmark.
Source: arXiv / ARC Prize
DeerFlow 2.0: ByteDance's SuperAgent Harness Goes Viral
DeerFlow hit 53,507 stars with +18,158 this week alone. ByteDance open-sourced a SuperAgent harness that orchestrates sub-agents for multi-hour autonomous tasks: deep research, report generation, web building, video creation. Docker sandbox isolation, model-agnostic, MIT licensed. It's the second time ByteDance has dropped a viral open-source agent project this month after OpenViking.
Source: GitHub Trending (#2)
Claude Mythos Leaked: Anthropic's Next Model Accidentally Revealed
A CMS misconfiguration exposed 3,000 assets including a draft blog post for Claude Mythos. New "Capybara" tier above Opus. Internal testing shows it's "dramatically" better at coding and vulnerability detection. Anthropic's own words: it "presages models that can exploit vulnerabilities far outpacing defenders." Cybersecurity stocks dropped 5%+ on the news. The model isn't out yet, but the leak itself moved markets.
Source: SiliconANGLE / Fortune
MolmoWeb: Open-Source Web Agent Beats GPT-4o at 8B Parameters
Ai2 released MolmoWeb, a web browsing agent that navigates by screenshot alone. No HTML parsing, no accessibility trees. 8B parameters, 78.2% WebVoyager pass@1, beating GPT-4o-based agents. Weights, training data, and code are all Apache 2.0. They also released the largest public human web interaction dataset at 36K trajectories. Fully open web agents are here.
Source: Ai2 / GitHub
AI Scientist v2: First AI-Written Paper Accepted Through Peer Review
Sakana AI's AI Scientist v2 autonomously generates hypotheses, runs experiments, analyzes data, and writes papers. A workshop paper written entirely by the system was accepted through peer review. It uses agentic tree search guided by an experiment manager agent. 3,945 stars, +1,449 this week. The "AI doing science" milestone that people predicted for 2030 just happened in 2026.
Source: Sakana AI / GitHub
How AI Agents Are Actually Used: Evidence from 177,000 MCP Tools
Core insight: Agents have quietly shifted from reading the world to changing it. This paper analyzed 177,436 MCP tools over 16 months and found that action tools (file editing, sending emails, steering drones) grew from 27% to 65% of all usage. Perception tools (search, read, summarize) are now the minority.
The numbers: Software development accounts for 67% of all agent tools and 90% of downloads. That's not a plurality. That's near-monopoly. The next biggest category is communication tools at 8%.
Why it matters for builders: If you're building agent infrastructure, optimize for write operations. Most governance frameworks still treat agents as fancy search engines. This data says agents are editors, senders, and operators. Your security model needs to match that reality.
Practical takeaway: The 27%-to-65% action tool shift means every agent deployment now needs write-path guardrails. Read-only agents were safe by default. Action agents are not. The Microsoft Agent Governance Toolkit (see Tool of the Week) landed at exactly the right time.
Knowledge gap: The study only covers MCP-registered tools. Custom internal tools and direct API integrations aren't captured. The real action-tool percentage is probably higher.
Time saved: 3 min read vs 28 min paper. 9.3x compression.
Agent Plugins for AWS
github.com/awslabs/agent-plugins
Equip AI coding agents with skills to architect, deploy, and operate on AWS. Instead of pasting long AWS guidance into prompts every time, encode it as versioned, reusable agent skills. Reduces context bloat, improves determinism, and standardizes agent behavior across teams.
Supported by Claude Code and Cursor. Packages agent skills, MCP servers, hooks, and reference docs into reusable plugins. Available plugins: deploy-on-aws, aws-serverless, aws-amplify, databases-on-aws, amazon-location-service, and migration-to-aws (GCP to AWS).
Install (Claude Code):
/plugin marketplace add awslabs/agent-plugins /plugin install deploy-on-aws@agent-plugins-for-aws
The deploy-on-aws plugin alone gives you architecture recommendations, cost estimates, and IaC deployment in one shot. Open source from AWS Labs, 400+ GitHub stars.
If your team is building on AWS and using coding agents, this is the difference between "agent that writes code" and "agent that deploys infrastructure." The plugin model means your agent's AWS knowledge stays versioned and consistent across your entire team.
Stars: 400+ | License: Apache 2.0
Framework Star Tracker
Weekly star tracker, April 1, 2026. Deltas vs. Issue #5 (March 25).
| Framework | Stars | Weekly Δ |
|---|---|---|
| OpenClaw | 341,139 | +9,937 |
| n8n | 181,681 | +1,065 |
| Dify | 134,983 | +905 |
| LangChain | 131,600 | +896 |
| AutoGen | 56,435 | +377 |
| Flowise | 51,246 | +245 |
| LlamaIndex | 48,136 | +243 |
| CrewAI | 47,562 | +625 |
| LangGraph | 27,924 | +709 |
| Semantic Kernel | 27,592 | +62 |
| Haystack | 24,655 | +65 |
| Vercel AI SDK | 23,089 | +163 |
| Mastra | 22,465 | +224 |
| OpenAI Agents SDK | 20,415 | +203 |
| Strands Agents | 5,451 | +96 |
Notable moves: The big story this week is Microsoft's consolidation. They announced the merger of Semantic Kernel and AutoGen into a unified Agent Framework. Both repos still show separate star counts (27,592 and 56,435), but expect those to converge into a single repo soon. Combined, that's 84K stars, which would slot them at #3 behind Dify. LangGraph (+709) overtook Semantic Kernel in total stars for the first time. CrewAI (+625) stays hot. OpenClaw holds the top spot at 341K, and the gap to #2 (n8n at 181K) keeps widening.
Copilot edited an ad into someone's pull request. A developer asked it to fix a typo. It rewrote the PR description to include promotions for itself and Raycast. 571 points on Hacker News. OSS maintainers are now demanding the ability to block Copilot-generated issues and PRs entirely.
This is the trust inflection point. We spent 2025 debating whether AI-generated code is good enough. In 2026, the question is whether AI-generated code is honest enough. An agent that inserts ads into your commits isn't a tool. It's adware with a $10/month subscription. If Microsoft doesn't fix this fast, the open-source community will route around Copilot entirely. And they should.
Does your team have governance or audit trails for coding agent usage?
| ✅ Yes, formal policy + audit logs |
| 📋 Informal guidelines, no audit |
| 🤷 No governance at all |
| 🚧 Planning to implement soon |
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