Most products labeled 'agentic OS' are orchestration libraries with a marketing budget. A working definition, drawn from the parts of the stack that actually have to behave like an operating system.
Twenty-odd products are calling themselves agentic workforce platforms. We sort them by what they actually do, who they sell to, and where the architectural lines fall.
A reading of four bundled agentic platforms — Sema4.ai, MultiOn, Adept's ACT family, and Web4OS — and the architectural patterns that recur across them: supervisor topologies, card-stream surfaces, credit schedulers, and canonical-host integrations.
Two protocol efforts are quietly defining how agents will talk to each other and to the rest of the world. A reading of what they are, what they leave open, and where the lines are still being drawn.
A working engineer's notebook on deploying OpenHands, the open-source coding agent that grew out of OpenDevin, against a real codebase. The runtime topology, the model adapters, the gotchas, and the production-readiness ledger.
The agentic-OS label gets diluted because it sounds important. It sounds important because it is. An argument for keeping the metaphor and what it actually demands.
Six years of agentic orchestration in one diagram. From cron-and-prompts to supervisor-led workforces, the path the field took and the steps still to come.
An engineer-eye comparison of the two major hosted agent surfaces from the model labs, including the rate limits, tool-execution semantics, and the production failure modes each one rewards and punishes.
Five products at five different layers of the agentic stack. A qualitative teardown across topology, surface, scope, and the kind of user each one is for.
A year after Anthropic published the Model Context Protocol, the spec has more clients than tools, more reference servers than production servers, and a clearer set of unfinished problems than most observers want to admit.
Identity is the problem the agentic field has spent the least time on and the one that determines whether the systems are deployable in any regulated environment. A working framework for what identity actually has to do.
A working-engineer's selection guide across three of the most-deployed open-source agent frameworks, against the public 2025 release notes and the production patterns each one rewards.
Per-seat pricing is wrong for agentic products. Per-call pricing is fragile. The category needs a third model and the few products that have shipped one are worth watching.
A working engineer's read on n8n's AI Agent and LLM nodes, the shape of agentic workflows you can build without a Python framework, and where the abstraction breaks down.
A working list of the agentic products doing architecturally interesting work in 2026. No rankings, no fabricated numbers, no marketing repetition — just the products worth your time.
Anthropic donated MCP. Google donated A2A. The substrate of agentic AI now lives under a neutrally-governed foundation. A reading of what that means architecturally, what it means politically, and what the next twelve months of standards work probably looks like.
Andrej Karpathy calls agentic AI 'slop.' Claude Code is at a $2.5B run-rate. Cursor is at $2B ARR. Devin's ARR went from $1M to $73M in nine months. Both readings are honest. The Review's attempt at why.
Four open-source coding agents — OpenHands, Aider, Continue, and the post-fork Cline lineage — now cover most of what a closed-source IDE assistant does. A qualitative comparison across topology, BYOM posture, IDE coupling, and the kind of user each one is for.
Three architectural framings are competing for the same agentic-systems problem. Each gets something right. Each papers over something important. A reading of the OS metaphor, Nate Jones's Mesh thesis, and the Workforce model that Anthropic and Sierra have leaned into.
Planner / Specialist / Tools
ReAct loop (single-agent)
MCP client / server topology
The Agentic Review is a technical publication covering the architecture, products, and standards of the agentic AI era. Long-form teardowns, comparative analysis, and protocol commentary, written by working engineers and architects.
We do not publish synthetic benchmarks. Comparisons are qualitative, based on published architectures and the experience of working practitioners. See our editorial guidelines for the full policy.