Opportunities › Agent behavior monitoring at scale › Proposed app
Tellwatch
Check every agent turn for looping, leaks and frustration without paying frontier-model…
- Score
- 46
- confidence 61
- Evidence
- 8
- items on the page
- Rivals
- 2
- named in sources
- First version
- 60 d
- estimate
- Type
- Desktop app
- PWA: runs in the browser, installs on Mac and Windows
- Market since
- 2023
- Emerging market · AI estimate
ScreensConcept
Desktop app: ML engineers review agent traces, classifiers, and failure patterns in consoles with dense data requiring sustained analysis and detailed configuration. Sample screens on a computer, in the proposed brand. Their numbers are sample data.
Behavior overview
Last 7 days1.8M turns
About this appAI proposal
Tellwatch is an API and web console that reads your AI agent traces and flags behavioral failures on every turn: looping, reasoning leakage, user frustration and more. Small classifiers run side by side on shared computation, so adding a signal adds little cost.
It is built for ML engineers and startups running production agents with tens of thousands of runs and millions of turns, who now either pay a frontier model to judge each turn or sample traces by hand and miss most failures.
You send turns through the API or a batch upload and get labels back. Alerts and a short review list in the console show which runs to open first. No dashboard to stare at: the labels go to your own pipeline, alerts and evals.
- Tagline
- Every turn checked. Nothing to stare at.
- Voice
- PlainPreciseCalm
- Why this name
- Tell is the small giveaway that something is off, as in a poker tell. Watch says it never stops. Two plain words in one.
In the first version
Members say which features they would use. It tells PWAS what people want; it changes no score.
Many classifiers, one pass
Run looping, reasoning leakage, frustration and your own signals together on a single read of each turn.
No member yet
Fast enough for every turn
Designed for sub-100ms responses so checks can cover 100% of turns rather than a sample.
No member yet
API first
Post a turn or a whole trace, get labels and scores as JSON. Webhooks send flagged runs to your own tools.
No member yet
Train your own signal
Label a few dozen examples from your traces and add a custom classifier for your own failure mode.
No member yet
Brand identityAI proposal
Check the name against trademark registers before using it.
Palette
#0F3D3EDeep Teal · primary#F5F2EAPaper · surface#17211FInk · text#E8A317Signal Amber · accent#B4472BAlert Rust · data, warnings#6B8F71Moss · data, healthy
Typography
Aa
Tellwatch
PricingAI proposal
The material gives no competitor prices. These tiers are a starting point to test, aimed to cost less than judging every turn with a frontier model.
Free
Free
- 50k turns a month
- 3 built-in classifiers
- Console and API
Team
$149 a month
- 2M turns a month
- All built-in classifiers
- Webhook alerts
- 2 custom classifiers
Scale
$799 a month
- 20M turns a month
- Unlimited custom classifiers
- Judge spot-check tools
- Priority support
Would you pay Team · $149 a month?
Your answer tells PWAS which prices people accept. It changes no score.
No answers yet.
What it could be worthEstimate
What a software business making this much sells for today, from real deals. Pick a monthly revenue.
If it makes
Worth about
$194,000
Sale range $132,000 to $263,000
$120,000 a year in revenue
From 87 software businesses listed and sold on Empire Flippers in the last year, revenue vetted by the broker: they sell for about 19× their monthly revenue, most between 13× and 26×.
Measured on profit, the middle deal is 32× monthly net profit: about 2.7 years of profit.
An estimate, not an offer or financial advice. A real price depends on profit, growth, churn, how much of the work is the owner's, and the buyer.
How it would growAI proposal
How the product would reach new customers through its own use, and how the first ones would be found. A plan to test, not a result.
- Built-in loop
- Partners and integrations. Ship SDK plugins and callbacks for common agent frameworks and tracing tools, so teams add Tellwatch to a trace pipeline they already run, and listings in those ecosystems send new users.
- Do they become customers?
- Integration users are engineers who can sign up and pay directly. Test whether they convert from the free tier once turn volume passes the free limit.
- First customers
- ML engineers at seed to Series B startups running customer-facing AI agents with over a million turns a month and a frontier-model judge bill.
- Where to reach them
- Hacker News and developer forums: Publish a cost and latency comparison of classifiers against a frontier judge on open trace data, and invite replies.
- Direct outreach: Message ML leads at startups that publicly describe running production agents and offer a free trace audit.
- First milestone
- Five startups send live traffic for 30 days and two convert to a paid Team plan.
Why people already pay for this
- VerifiedHacker News post received 20 points and 2 comments
- Public claimProduction agents have common behavioral failures like looping, reasoning leakage, and user frustration that need monitoring.
- Public claimUsing frontier models to judge every agent turn is too expensive and slow at scale.
- AI inferenceMedium-sized startups deal with tens of thousands of agent runs and millions of turns requiring behavioral tracking.
- AI inference
Tech stack
Every proposal builds on the same PWA base, installable from the browser on any phone or computer. Dashed boxes are what this app adds.
- AppNext.js and TypeScript
- PWAManifest and service worker
- UITailwind CSS
- Data and authSupabase Postgres
- PaymentsStripe
- EmailResend
- HostingVercel
- This app · Classifier inferenceGPU or CPU inference service with batchingShared computation across many small classifiers is the core of the speed and cost claim.
- This app · Trace ingestionQueue and columnar store for high-volume turnsMillions of turns outgrow plain Postgres for storage and aggregate queries.
RisksAI inference
- TechnicalReaching sub-100ms latency with shared computation across many classifiers is unproven here and may need real research before it holds.
- AISmall classifiers may be less accurate than frontier judges on subtle failures, and may drift as agents change.
- CommercialEvidence is a single Hacker News post with 20 points and 2 comments; demand is unverified, and teams may build in-house.
- DataAgent traces hold sensitive user data, so customers may demand strict retention, redaction or self-hosting.
- PlatformObservability vendors or model providers could add cheap behavioral checks to their own products.
Everything marked AI proposal or AI inference was written by a model from the evidence on the opportunity page. The proposed name must be checked against trademark registers before it is used.