What changed
FACT (per cited signals): Google shipped computer use in Gemini 3.5 Flash, a low-cost low-latency tier for UI-driving agents (deepmind.google blog). A stealth Chromium exposed over MCP (Fortress, tilion.dev) launched so agents can browse sites that block automated traffic. HYPOTHESIS: together these collapse per-task cost AND the 'agent gets blocked' failure mode, making fixed-price done-for-you automations honorable margins for a solo operator.
Why now
Per-task inference cost for browser agents just dropped an order of magnitude (Flash-tier computer use), and the single biggest reliability killer — bot detection — now has an off-the-shelf MCP-packaged answer. Before this pair, flat-fee automation contracts were margin roulette; a solo builder could not price them safely. HYPOTHESIS: a 6-12 month window before incumbents (Apify, Browserbase, Zyte) bundle the same stack as a commodity.
Converging signals
(1) Computer use in Gemini 3.5 Flash — cheap agentic screen/browser control [https://deepmind.google/blog/introducing-computer-use-in-gemini-3-5-flash/]. (2) Fortress — engine-level fingerprint-spoofed Chromium over MCP so agents don't get blocked [https://tilion.dev]. Convergence is real at the CAPABILITY layer; nothing in the input evidences DEMAND.
Customer pain
HYPOTHESIS ONLY: no demand_evidence array was supplied (it is empty). The asserted pain — 'businesses need repetitive data pulled/filed from sites that block bots' — is plausible and adjacent to known markets (scraping, RPA), but this input contains zero complaints, job posts, or mandates. Per the system's own lesson (conf 0.85), demand is the scoring bottleneck and must not be invented.
Who pays
HYPOTHESIS: SMB operators who today pay VAs or scraping-API credits — e-commerce sellers monitoring competitor price/inventory on hostile marketplaces, real-estate/lead shops pulling from listing portals, agencies doing recurring competitive intel. Reachable via cold outreach and niche communities. No evidence any of them will pay a flat fee is present in this input.
Solved today
Scraping APIs (Apify, Bright Data, Zyte, Browserbase), RPA (UiPath/Make/Zapier), offshore VAs, or in-house Playwright scripts that break when anti-bot changes.
Why current solutions are bad
DIY scripts break on anti-bot updates and need babysitting; scraping APIs are usage-metered (unpredictable cost) and still get blocked on the hardest sites; VAs are slow and error-prone. HYPOTHESIS: a managed flat-fee 'we keep it working' service removes both the cost-variance and the maintenance burden.
Proposed product
A managed automation service: customer describes a recurring web task (monitor X, file Y, extract Z); you build and host the agent on the Flash+stealth stack and charge a fixed monthly fee per automation, with an SLA on freshness/success. Thin dashboard for outputs + alerts; delivery via CSV/webhook/email.
MVP version
Pick ONE vertical with a concrete, legal, willing buyer (e.g. price/inventory monitoring for Shopify/Amazon resellers). Hand-build 3-5 automations on the Flash computer-use + stealth-MCP stack, run them on cron, deliver a daily diff report. Sell as a $99-299/mo managed service. Do it as a service first, productize the repeated 20% later.
30-day build
Validate demand BEFORE building: 30-40 customer interviews / cold DMs in one vertical to confirm willingness to pay a flat fee and the specific site targets. In parallel, stand up the Flash computer-use + Fortress MCP stack and prove reliability on 2-3 real target sites. Kill if you cannot get 5 verbal 'yes I'd pay' in 30 days.
60-day build
Onboard 3-5 paying design-partner customers at a founder price. Build the thin ops layer: scheduling, success/failure alerting, retry, output delivery. Codify a legal posture (only automate sites where the customer has a right/relationship, or public data; get customer indemnity in ToS).
90-day revenue plan
10-20 flat-fee automations live at $99-299/mo → ~$1.5-5k MRR. Standardize onboarding to <1 day per automation; begin templating the top-requested tasks into near-self-serve.
Distribution path
Demonstrated-value cold outreach (record a Loom running THEIR automation live), niche subreddits/Slacks/Discords per vertical, and 'I'll automate your worst recurring task free once' hooks. No ad spend needed.
Pricing hypothesis
Flat $99-299/mo per automation; setup fee $199-499 to filter tire-kickers and fund the build. Higher tiers for higher freshness/volume. Avoid usage metering — the flat fee IS the differentiation.
Technical difficulty
Moderate. The stack exists (Gemini Flash computer use + Fortress MCP); the hard part is reliability/maintenance ops and keeping automations alive as sites change — an ongoing labor cost, not a one-time build.
Legal / regulatory risk
HIGH and the core concern. The wedge is explicitly 'browse sites that block automated traffic' via fingerprint spoofing — that is bypassing access controls / violating ToS on hostile sites, with CFAA (hiQ v. LinkedIn nuance), Computer Fraud, and platform-abuse exposure. Defensible ONLY if narrowed to public data or sites the customer is authorized to access. As pitched (hostile-site evasion) it is a legal and reputational liability.
Platform dependency
High double dependency: on Google's Gemini Flash computer-use API (pricing/availability/policy) AND on Fortress/tilion.dev (a brand-new single-vendor MCP stealth product that anti-bot vendors will actively target). Both can change terms or break overnight.
Founder fit
Moderate. Strong on automation/AI workflows and demonstrated-value selling. BUT this is NOT the founder's proven government-portal forced-buyer shape (his highest-fit pattern per lesson conf 0.80) — there is no mandate, no forced buyer, no per-filing regulatory hook. It's a generic managed-automation service in a crowded, legally gray market.
Breakout potential
Moderate. Could grow into a templated micro-SaaS if a repeatable vertical emerges, but it sits in a commoditizing space (scraping/RPA) where incumbents with capital can bundle the same Flash+stealth stack fast. Not a durable moat.
Final recommendation
REVISIT-LATER / CONDITIONAL. Do not build as pitched (hostile-site evasion). The capability convergence is real, but this input carries no demand evidence and the wedge is legally gray and easily copied. IF pursued, pivot hard to a narrow, legal, willing-buyer vertical (authorized/public-data monitoring) and validate willingness-to-pay first — or better, apply the same Flash+stealth stack to the founder's proven government-portal forced-buyer pattern, where demand is structural and fit is highest.
Next action
Spend one week testing demand in a single legal vertical (e.g. 20 cold DMs to Amazon/Shopify resellers offering free price-monitoring). Require 5 concrete 'yes I'd pay a flat fee' before writing production code. If none, shelve and redeploy the stack toward a mandate/forced-buyer opportunity.