Agentic runtimes, autonomous engines and the models underneath them - indexed, costed per task, and wired into systems that run.
The same daily workload costs $30 or $1,620 a month depending on one routing decision. We publish which one, and the setup that gets you there.
Free. One weekly email. Unsubscribe in one click.
Everything else drowns you in launches. The Agentic Radar reads all of it and sends you the handful of lines that change what you should build, run, or route this week - nothing else.
A curated brief you can use, not another feed to skim. Built to scan in under ninety seconds, grouped by what changed rather than who shipped it, and interactive - filter down to the categories that touch your stack and watch the rest collapse out of the way.
Weekly. Unsubscribe in one click. Subscribers get founding access to everything below.
An agent's bill is not a subscription. It is tokens in, tokens out, multiplied by how many times a day it runs - plus whatever the runtime costs to stay awake. Pick the task size first: that is where every cost estimate on the internet goes wrong.
Rates are public API list prices per million tokens, reviewed 25 August 2026. Token sizes per task class are our own measured working assumptions, not vendor figures - they are the number most calculators hide. The thinking uplift is applied to the frontier tier only, where extended reasoning runs by default and bills at the output rate; other tiers are costed without it. Caching assumes roughly 70% of input served from cache at 10% of the input rate. The heartbeat assumes 48 wakes a day at about 8K context each. Runtime fees are published vendor tiers.
One task here means: the agent reads roughly fifteen thousand tokens of context - five web pages, a thread, or a handful of files - reasons over it, and writes back about fifteen hundred tokens. That is a real unit of work, not a chat message. Drag the volume and watch the same work reprice.
15,000 tokens in, 1,500 out, 30 days, thinking included on the frontier tier.
List prices per million tokens, verified 25 August 2026. The gap between the top and bottom bar is one configuration line - and it is worth more than every subscription you could cancel this year.
Correctly wired means: task classes separated, each routed to the model tier that can actually do it, caching on, triggers instead of polling, and a human on the exceptions. Wired that way, these are the numbers it plays against.
Per month at the top of the US full-time virtual assistant range - $800–2,200 from India, $1,100–2,200 from the Philippines. A $76,590 executive assistant salary loads to roughly $109,000 a year once benefits and payroll tax are counted. A wired agent covering the same routine volume runs at two to three figures a month, not five.
Sources: 2026 VA pricing surveys, Glassdoor salary dataA full-time support hire, salary and benefits, per year. The agent covering the routine tier of the same queue runs $1,200–2,400 a year.
Source: small-business agent cost analysis, June 2026Per month for an AI receptionist handling 85–95% of inbound calls, against roughly $45,000 a year for the human equivalent.
Source: small-business agent cost analysis, June 2026The share of routine assistant work agents carry autonomously in 2026 - inbox, scheduling, research, drafting, reporting. The remaining slice is judgment and escalation, and every setup we publish assumes a human owns it. That assumption is the difference between a system that compounds and one that gets switched off in month two.
Source: solopreneur agent-stack analysis, March 2026The strongest open models now sit at the top of independent indexes under MIT and Apache licences, at a fraction of frontier token prices. Knowing that changes nothing on its own - the value is in the setup: which task classes move, what the runtime looks like, what breaks, and what it costs at your volume. That is what we publish, per model.
For each leading open model: the runtime that hosts it well, the task classes it handles at frontier-equivalent quality, the ones it does not, and the failover path back to a frontier tier.
In buildTask class to model tier, with the price per task written next to it. The single artifact that turns a fifty-times spread into a line of configuration.
In buildWhat each open model actually needs to run - hardware, quantisation trade-off, hosted-API alternative - and the point where self-hosting stops saving money.
QueuedWe name models where the price matters. We do not sell them, and we do not send you off to compare leaderboards - the leaderboard is not the decision. The wiring is.
Five products in the order you need them. Each one ends with a number - what it saves, or what it earns.
Weekly monitoring with a fixed structure - what moved, whether it touches your setup, the action, and the share line. Interactive, semantically categorised, ninety seconds to scan.
A short quiz - what you run, what you want autonomous, your budget, who maintains it. Out comes the named runtime, the routing per task class, and the tools you do not need to buy.
The constructor: your Stack answers turned into the build itself - quiz to configuration to specification. Workflow, exact configuration, permissions, memory, run order, and what to consider at each step, and the monthly cost at your real volume - written before you build, not after.
Complete wirings that already earn: components, connections, run order, and the monthly cost at real volume. You are copying a working setup, not a diagram of one.
Every agentic runtime, AI automation platform and programmatic framework on one schema - 10 to 20 filters per category, what it costs at your volume, where it breaks, and how to wire it. Filter by constraint, not by category name.
A five-step workflow and a full autonomous organisation are not the same document. Every specification is quoted against your system's nodes, steps, integrations and complexity - there is no flat rate, because there is no flat build.
Every flow needs different steps, setups and wiring, so the price is set on a discovery call - nothing is charged before that. The Early Access discount is locked the moment you subscribe.
Confirm the email to hold your place, then reply with what you run.
Four phases. Each one opens to Radar subscribers before it opens publicly.
The weekly brief and the free cost tools on this page.
Quiz to a named, routed setup you can act on the same day.
Build documents and complete wirings that already earn.
The hub: every agentic platform, automation runtime and autonomous engine on one schema.
The Agentic Radar ships weekly, free. Founding access to everything else comes with it.
We sent a confirmation link to your email. Click it and the Agentic Radar starts with the next issue — founding access unlocks the moment you confirm.
No email in a couple of minutes? Check spam — it lands under Automationery.