RXed AI News

What it is — in one coffee-break

An agent is a loop. The model thinks (picks a next step), acts (calls a tool, writes a file, hits your CRM), then observes what came back and decides again. Repeat until the goal is met or the budget runs out. That loop is the entire difference between a chatbot, which answers and stops, and an agent, which keeps going until the refund is issued or the pull request is open.

Three things separate a real agent from a demo: autonomy (how many steps it takes before it needs you), reliability (how often those steps end in the right outcome), and integration depth (whether it can actually touch your systems or only talk about them). A vendor can be strong on one and useless on the other two. Gartner counted roughly 130 genuinely agentic vendors out of thousands claiming the label, and named the rest “agent washing”.

When you actually need it (and when you don't)

You need agents when the work is multi-step, repetitive, high-volume and has a checkable definition of done: refund a customer within policy, triage 400 inbound tickets, chase one document across four systems, reproduce a bug and open a fix. The tell is that a human does it today by tabbing between five tools for twenty minutes.

You do not need agents for one-shot generation — write this email is a prompt, not an agent. You do not need them for deterministic work with fixed rules; a script or an n8n workflow is cheaper and never hallucinates. And you should not point them at irreversible actions with no undo: wire transfers, bulk deletion, sending to a full mailing list. Gartner expects over 40% of agentic projects to be scrapped by the end of 2027, mostly for escalating cost, unclear value and missing risk controls. A large share of those are teams who pointed an agent at a problem a rules engine had already solved.

How to recognize good vs bad implementations

Good implementations are boring in the demo and impressive in the logs. Every run is traceable step by step, failures halt instead of improvising, permissions are scoped per tool, and there is an explicit approval gate on anything expensive. Bad ones show you a flawless three-minute happy path and change the subject when you ask about run number 4,000.

Make the vendor answer three questions live: “Show me a run that failed — what did the agent do next, and where is that written down?” “How many tool calls deep does it go before a human has to step in, and what is the success rate at that depth?” “Which actions can it take without asking, and how do I reverse one after the fact?” A vendor who can only answer the first is selling a chatbot with ambition.

Calibrate with real numbers rather than the pitch. On TheAgentCompany, a benchmark of ordinary office work, the most competitive agent completed 30% of tasks autonomously. METR measures a 50%-success time horizon of roughly 320 minutes — about five hours of software work — for the strongest model it has tested, with that horizon doubling every three to four months. Long, open-ended tasks still fail. Short, scoped, well-instrumented ones increasingly do not.

What this costs

Three pricing shapes. Outcome-based: Sierra bills per resolved conversation rather than per seat, so unresolved chats cost nothing — honest, but your bill scales with volume forever. Consumption-based: Devin starts at $20 and burns $2.25 per ACU, where one ACU is about fifteen minutes of agent work; the old flat plan was $500 a month. Seat-based platforms: Glean contracts cluster around a median near $99,000 a year in Vendr's buyer data, ranging from about $30,000 to $209,000.

The free-and-on-premise route is real and it is what most readers here should try first. n8n's self-hosted Community Edition is free (their managed cloud starts at €20/month), and LangGraph, CrewAI and the major vendor agent SDKs are open source. Your hard cost becomes a VPS at €10–40 a month plus model tokens — realistically €50–300 a month for a small-business workload, or close to zero if you run an open-weight model on a GPU you already own. You pay instead in engineering time: the loop itself is a weekend, the evaluation harness, retries, permissions and audit log are the six weeks.

Where to see it scored

We audited 62 tools and scored 56 of them on Ag. Top of the table: Sierra (9.5) and Claude / Anthropic (9.5), followed by Windsurf, Devin Desktop (9) and Glean (9). Bottom of the table: Kling AI (3) and Cognism (3) — both competent products that simply do not run autonomous loops, whatever their marketing pages imply.