Always-On Agent

What is it

An Always-On Agent is an AI agent that stays persistently available — reachable at any moment, across channels, holding context between interactions — and that you pay for only while it is actually doing work. The technical essence is a long-running orchestration layer: a persistent memory store, a scheduler/trigger system (webhooks, cron, inbound messages), a sandboxed execution environment, and a metering layer that counts active compute seconds rather than seats or messages.

The business significance is bigger than the tech. Every incumbent AI product today charges the wrong unit. ChatGPT Plus charges $20/month whether you use it ten minutes or ten hours. Agent platforms like Lindy and Zapier charge task credits that are opaque and hard to forecast. Always-On Agents flip the model: the agent is there when you need it, and the invoice tracks actual work. That reframing does two things — it removes the buyer's fear of paying for idle capacity, and it aligns vendor revenue with delivered value. For indie developers, this is a wedge into a market currently owned by seat-based pricing that nobody loves.

Why now

Three things converged in late 2025 and early 2026. First, inference costs for small-to-mid models dropped sharply — running a persistent agent loop on a cheap model with escalation to a frontier model on hard steps is now economically viable at consumer prices. Second, tool-calling and computer-use APIs matured to the point where an agent can reliably take actions (send email, file tickets, update CRM) rather than just chat. Third, the pricing conversation shifted: developers on Show HN and Product Hunt started publicly questioning credit-based agent pricing, and products like Rex and Botel began testing "billed only while working" as a differentiator.

The demand signal is specifically about predictability. Users are tired of either flat subscriptions they underuse or credit systems they cannot forecast. "Always reachable, billed only when working" is a legible promise. This is also why it is emerging now rather than last year: last year the models were not reliable enough for persistent autonomy, so an always-on agent would have burned money while failing. The reliability floor has moved.

Market Evidence

The signal is early but coherent: 3 independent sources (Product Hunt, V2EX, Show HN), 3 total mentions, 100% growth rate, stage classified as nascent, trend score 72/100. Three mentions is thin — treat this as a leading indicator, not a proven market. But the distribution of sources matters: Product Hunt gives you product-launch signal, Show HN gives you builder signal, and V2EX gives you practitioner signal from a technically sophisticated, cost-sensitive community. When the same interaction model (always reachable, pay-per-work) shows up in all three, it is more than one founder's idea.

What is missing is the demand-side proof: no visible waitlists, no revenue disclosures, no complaint threads at scale. That is exactly what you would expect at 3 mentions. The honest read: this is a real and rising conversation about pricing and interaction model, not yet a demonstrated market with buyers lining up. The window to build is now, while competitors like Rex and Botel are still defining the category and before a platform incumbent standardizes the meter.

Who's Behind It

The named players are Rex and Botel, both early-stage and both experimenting with the always-on interaction model and work-based billing. Neither has meaningful brand gravity yet — they are category definers, not incumbents. Behind them sit the platform whales whose pricing decisions will determine whether this stays a niche: OpenAI (Assistants/Agents API), Anthropic (agent tooling), Google (Gemini agent APIs), and the workflow-automation tier — Zapier, Lindy, Relevance AI, and n8n. The platform layer is the real competitive risk; the named startups are the proof of concept.

The community driver is the indie builder crowd on Show HN and Product Hunt, plus the cost-conscious practitioner base on V2EX. That community is small but high-conviction and vocal, which is why three mentions produced a trend score of 72. Watch the whales' pricing pages, not their blog posts — the moment one of them ships metered "active compute" billing, the differentiation window narrows sharply.

TAM & Market Size

The buyer is not "everyone who uses ChatGPT." It is the person who already pays for automation: the solo founder running a one-person agency, the ops lead at a 10-50 person company, the developer who wants an agent on a webhook. Start with the SMB and prosumer automation segment — Zapier alone reports millions of users, and the subset who want persistent, context-holding agents rather than trigger-action zaps is a credible beachhead.

Price tolerance: the automation buyer already pays $20-$100/month for tools (Zapier Professional starts around $30/month, Lindy tiers run $50-$200+). A work-metered agent at a $0 base fee plus roughly $0.10-$0.50 per active minute, capped by a $49-$99 monthly ceiling, fits inside existing budgets without a new procurement conversation. The provided scores — opportunity 0/100, demand 0/100, market 0/100 — are placeholder zeros from an unscored record, not a verdict; the underlying trend score of 72 and 100% growth are the real signals. Treat TAM as unproven but plausible: thousands of early adopters, not millions, in year one.

Competitive Landscape

Direct competitors: Rex and Botel, both nascent, both pursuing the same interaction model. Their weakness is that they are building the whole stack — memory, execution, billing — as small teams, which means shallow integrations. The workflow tier (Zapier, Lindy, n8n, Relevance AI) has distribution and integrations but seat/credit pricing that the always-on thesis directly attacks. The platform tier (OpenAI, Anthropic, Google) has the models and could ship metered agent billing at any time.

The gap: nobody owns transparent work metering as a product. A developer-facing API that any agent builder can drop in to bill their own users by active compute — Stripe for agent time — is an unclaimed position. Competition score 0/100 reflects an unscored field, not an empty one. If Big Tech enters with native metered billing, you likely have 9-12 months of differentiation; the defense is being the neutral, cross-model metering layer rather than another agent app.

Business Model

Recommended: usage-based SaaS with a hard monthly cap, plus a free tier of roughly 60 active minutes per month. Charge $0.20 per active minute, capped at $79/month on the Pro plan, and $249/month for teams with shared agents and audit logs. Why this fits: the entire promise is "pay only while working," so a flat subscription contradicts the pitch, while pure metered billing without a cap scares buyers. The cap is the trust device — it makes the invoice forecastable, which is the actual complaint driving the trend.

12-month forecast, solo founder: conservative $3K MRR (roughly 60 paying accounts at ~$50 blended), base $12K MRR (~200 accounts, some team plans), optimistic $35K MRR (~500 accounts plus API resellers). CAC via Show HN, Product Hunt, and developer content: $30-$80 per paying account, with payback under two months at a $50 blended ARPU — attractive because the audience self-selects from launch channels. The risk in this model is metering disputes; publish the meter definition and give users a live usage dashboard from day one.

MVP Blueprint

Ship in 5-7 days. Core loop only: (1) a persistent agent with a memory store keyed to a user ID, (2) three triggers — inbound email, webhook, and a scheduled cron, (3) a sandboxed tool set limited to send-email, HTTP request, and a key-value store, (4) a metering layer that counts active seconds from task start to task completion and writes an immutable usage log, (5) a dashboard showing live meter and monthly cap. Cut multi-agent orchestration, cut a marketplace, cut voice.

Stack: Next.js on Vercel for the dashboard, Postgres (Neon or Supabase) for memory and usage logs, a queue (Upstash QStash or a simple Postgres-backed worker) for triggers, and the OpenAI or Anthropic API for inference with a cheap model as default and escalation on failure. Metering is the moat — build it as a standalone service with a clean API so it can later be sold on its own. Ship the API alongside the UI; the suggested product types are SaaS, Tool, and API, and the API is what gets you distribution through other builders.

Commercial Opportunities

1. Metered agent infrastructure API. Sell the metering-and-orchestration layer to other agent builders who want work-based billing without building it. Target: indie agent startups on Product Hunt. Expected $5K-$20K MRR within a year if two or three builders adopt. Beats building another consumer agent because you sell to builders who already have distribution.

2. Always-on inbox agent for solo consultants. An agent that watches a business inbox, drafts and sends routine replies, books meetings, and escalates the rest — billed per active minute. Target: freelancers and one-person agencies paying $30-$80/month. Expected $3K-$10K MRR. Beats a general assistant because the inbox is a concrete, daily, measurable workflow.

3. White-label always-on agent for niche verticals (e.g., property management maintenance requests). Sell to small vertical SaaS vendors who want an agent feature without building it. $500-$2,000 per license per month. Higher ACV, slower sales cycle.

Product Ideas

🥇 Meterly — "Stripe for agent time." A drop-in API that lets any agent product meter active compute and bill users by the minute with a hard cap. Target: indie agent builders. Why now: every competitor is hand-rolling metering badly, and the pricing debate is the trend's core.

🥈 Standby — an always-reachable personal agent on email and Slack that holds context for weeks and bills only for active minutes, capped at $49/month. Target: solo founders and consultants. Why now: Rex and Botel prove the interaction model but have no pricing trust story; a hard cap is the differentiator.

🥉 Nightshift — an always-on agent for developer workflows: watches CI, triages failing builds, opens fix PRs, and reports in the morning. Metered per active minute. Target: small engineering teams. Why now: developers are the most price-literate buyers and the fastest to adopt metered billing.

SEO Opportunity

Search interest is early — terms like "pay per use AI agent" and "AI agent billing" are rising from a low base, and SEO difficulty is effectively near zero for the long tail. Target five phrases: "always on AI agent," "AI agent pay per minute," "AI agent usage based pricing," "metered AI agent billing," and "AI agent that only charges when working." Competition is thin because incumbents rank for "AI agent" head terms. Strategy: publish a transparent pricing comparison and a live cost calculator — calculators earn links and convert because they answer the exact question buyers are asking.

Risk Assessment

The thesis breaks if the always-on interaction model turns out to be a feature, not a product — i.e., if OpenAI or Anthropic ships persistent agents with native metered billing and users never seek a third party. Second risk: metering disputes destroy trust; users perceive "active minutes" as a rigged meter, and the product dies on Reddit. Third: reliability — an always-on agent that misfires at 2 a.m. costs the user real money and you the account.

Validate cheaply in two weeks: build the metering layer only, put a landing page with a live cost calculator on Show HN, and measure sign-ups for a waitlist. If fewer than 100 qualified sign-ups from a decent launch, walk away. Also run five customer interviews with Zapier power users about their current bill and whether they would switch to metered. If the answer is "I just want a flat price," the thesis is wrong.

Action Plan

Today: write the meter definition — exactly what counts as an active minute — and publish it as a one-page spec. That document is your product thesis and your first piece of content.

Week 1: ship a landing page with the live cost calculator and a waitlist, launch on Show HN and Product Hunt, and collect 100+ emails. Month 1: build the metering API plus a minimal agent demo, onboard five design partners on free metered access, and instrument their usage. Month 3: convert design partners to paid, publish real usage data as a case study, and decide whether to pursue the infrastructure API (Meterly) or the consumer agent (Standby) based on which cohort converts faster. Kill criteria: no paying customer by end of month 3.

Related Terms

Agent billing and metering — the direct sibling trend; the pricing debate is what makes always-on agents commercially interesting. Persistent memory agents — the technical enabler; without cross-session memory, always-on is just a chatbot with a cron job. Computer-use agents — the capability layer that lets an always-on agent actually do work rather than just talk, which is what justifies per-minute billing.