SuperAGI

The record system treats agents as actors rather than assistants, which is the architectural distinction this axis exists to catch.

Most platforms in this index bolt inference onto deterministic software. This one was built the other way round. It originates from an open source autonomous agent framework with substantial developer adoption, and the commercial product carries that lineage: the record system is described as tracking actions taken by both representatives and agents, meaning an agent is a participant in the pipeline rather than a tool a person operates.

The surface area is correspondingly wide. More than twenty five applications and agents. Agents that find and engage prospects continuously. Voice agents that hold telephone conversations, metered by the minute. A signals engine that resolves identity, enriches, scores and reasons about engagement without a person in the loop. Predictive scoring across a reported hundred variables. Reinforcement learning adjusting behaviour from outcomes.

The commercial model confirms it. Billing is metered in credits consumed by actions, and the published burn rates are for inference driven work rather than for storage or seats.

Strip the agents out and nothing remains that a customer would buy, because the proposition is explicitly replacing a stack of conventional tools with autonomous ones.

Ask what proportion of pipeline actions are taken by agents rather than people.

The platform is built around agents acting continuously and nothing describes what constrains them.

This is the axis an agentic product must answer hardest, and two passes across independent coverage, review analysis and every retrievable surface located no approval gate, no human in the loop configuration, no audit trail description, no stop condition, no escalation path and no permission model governing who may deploy an agent.

What is described is the opposite: agents working around the clock to find and engage prospects without hiring, a record system logging actions taken by agents as first class events, and voice agents holding telephone conversations billed by the minute.

One partial control appears in third party summaries of the upper tier, covering role based access, which governs who may use the platform rather than what an agent may do once running.

The reinforcement learning claim sharpens the question rather than answering it. A system adjusting agent behaviour from live outcomes is changing what it does over time, and nothing describes what bounds that adjustment or who reviews it.

Compared against a record built earlier this session where every model action required a representative's confirmation before saving, the contrast is complete.

Ask what an agent cannot do without human approval, and who can stop one mid conversation.

A platform whose entire identity is model driven, disclosing no model.

Two passes across independent coverage, review platform analysis and both vendor domains located no model provider, no model name, no version, no statement of what customer records, conversation transcripts or prospect data are transmitted for inference, no training or retention position and no evaluation.

The scale of what is unexplained is the point. A context graph reported to process more than one and a half billion data points daily. Predictive scoring across a hundred variables per record. Engagement reasoning inside the signals engine. Voice agents conducting conversations. Reinforcement learning altering behaviour from outcomes.

One accuracy figure exists and it concerns data rather than model behaviour: enrichment is claimed at ninety five percent accuracy through a waterfall approach.

The open source origin makes the silence more conspicuous rather than less. A company whose framework is public, forked widely and inspectable has an unusual opportunity to be specific about what runs underneath, and the commercial product publishes nothing.

Credits are consumed per action at published rates, so a buyer can see what inference costs without being told what performs it.

Ask which models run the voice agents and the engagement reasoning.

Four hundred and forty one reviews with a published distribution, and one consolidation claim specific enough to check.

The rating sits at 4.6 across 441 reviews, with the breakdown published as 342 at five stars, 86 at four, 11 at three, none at two and 2 at one. A distribution that skewed is worth reading carefully, and the volume is large enough to be meaningful.

The review themes are quantified by mention count, which is unusually granular: ease of use at 273 mentions, interface at 181, support at 179, automation at 105 on the positive side; missing features at 72, bugs and technical issues at roughly 89, limited features at 39 on the negative.

Five named customers appear across independent coverage, spanning developer tooling, communications, collaboration, healthcare and security. One is reported to have retired eleven separate go to market tools on adopting this platform, which is the single most useful data point available because it is a specific, falsifiable consolidation claim rather than a percentage.

Funding is reported at roughly fifteen million dollars including a ten million round with a named lead investor.

The deduction is the absence of outcome figures. No named customer reports pipeline, meetings or revenue, and the vendor's own surfaces could not be read.

Ask which customers have published measured outcomes.

A named customer states in a detailed review that automated unsubscribe handling was not in place.

The account is specific and worth setting out with its weight. The reviewer describes entering an annual agreement, being unable to send a single campaign for months owing to defects, data problems and misconfiguration, and identifies the absence of automated unsubscribe handling as introducing significant regulatory risk on an email marketing platform. They report escalating to the founders to terminate and recover a refund.

That is one review among four hundred and forty one, and only two carry the lowest rating, so it is plainly not representative of the general experience. It is recorded because the allegation is specific rather than vague, concerns a legally required mechanism rather than a preference, and no counter evidence was located on any retrievable surface.

Around it, multiple independent sources agree on three further absences: no warm up engine, no domain rotation and no address control for teams running cold outreach through the agents.

One third party listing claims compliance with the European data protection regulation, a service organisation control examination and health information privacy rules, which if accurate is a strong position and could not be verified here.

Ask whether automated unsubscribe handling ships by default on every sending channel.

Three compliance regimes claimed by a third party, and one capability that raises the sharpest privacy question in this batch.

The claim, from an independent listing rather than a vendor page, is that the platform meets the European data protection regulation, a service organisation control examination and health information privacy rules, and is therefore suitable for regulated industries. Health information handling in particular would imply specific controls, and one named customer operates in healthcare, which makes the claim plausible. None of it could be verified in this pass because neither vendor surface was readable.

The capability worth naming is anonymous visitor identification, described as resolving which companies and which individuals are browsing a customer's website in real time. Identifying a named person from an unidentified website session requires matching against an identity graph assembled from sources the visitor never engaged with, and it is among the most privacy sensitive functions anywhere in this index.

The holdings compound it: a database reported at over two hundred and seventy five million contacts, conversation transcripts, behavioural signals and a graph processing over a billion data points daily.

Two passes located no processing agreement, subprocessor list, retention schedule or residency statement.

Ask how individual website visitors are identified and on what basis.

Two different record counts circulate for the same database and neither carries an origin.

Independent coverage gives the contact database as more than two hundred and seventy five million records alongside seventy five million companies in one source, and more than four hundred and fifty million leads in another. Both are recent and both attribute the figure to the vendor. A buyer cannot tell which is current, and the gap is large enough to matter.

Enrichment is described as waterfall, meaning multiple sources are queried in sequence until a value is found, with accuracy claimed at ninety five percent. That method inherently involves several upstream suppliers and not one is named.

Two passes located no supplier, no collection method, no consent basis and no refresh cadence for any of it.

The identity resolution behind anonymous visitor identification is the sharpest unanswered question, since matching a website session to a named individual requires a persistent identity graph whose construction is nowhere described.

What is worth crediting: the prospecting, signals and visitor identification capabilities are reported to carry no additional charge on top of the subscription, which is unusual for data of this kind and means a buyer is not separately metered for lookups.

Ask which figure is current and which suppliers the waterfall queries.

No evasion language anywhere, and two capabilities whose basis is unestablished.

The clean part first, and it is worth stating because two other vendors in this same batch failed it badly. Two passes located no undetectability marketing, no claims about staying under the radar, no randomised limits sold as avoiding detection and no proxy infrastructure framed as evading platform security. For a product this aggressive in its automation, the absence of that language is a deliberate difference.

Two capabilities still need a basis. A plugin for the professional network is named in independent coverage, with no statement of what it does or under what agreement. And messaging platform outreach appears among the channels, carrying opt in and template approval requirements that nothing addresses.

Anonymous visitor identification sits at the edge of this axis. Resolving a website session to a named individual depends on an identity graph built from data collected elsewhere, and whether that collection was licensed is unestablished.

Email, telephone, text and web push all run through sanctioned infrastructure.

The honest position is limited exposure with two open questions, rather than the deliberate boundary pushing recorded elsewhere in this batch.

Ask what the professional network plugin does and under what agreement.

Agents converse with prospects by voice and nothing published constrains, measures or governs them.

On safety, two passes located no model disclosure, no evaluation, no accuracy measure, no error rate, no guardrail description, no statement of what an agent may not say and no stop condition. Voice agents hold telephone conversations at a published rate of fifteen credits per minute, which prices the capability precisely while describing its behaviour not at all.

The reinforcement learning claim is the one that needs an answer most. A system that adjusts agent behaviour from live outcomes is modifying what it says to real people over time, learning from what worked, and no boundary, review process or rollback mechanism is described anywhere.

On stewardship, a third party listing claims a service organisation control examination and health information privacy compliance, which would be substantial if verified. Neither vendor surface could be read, so no encryption statement, access control description, incident response process, breach notification commitment, named security contact or vulnerability disclosure route is recorded here.

The custody is significant: a contact database in the hundreds of millions, conversation transcripts, and a graph processing over a billion data points daily.

Ask what constrains agent behaviour and what the reinforcement learning optimises for.

A synthetic voice holds a telephone conversation and nothing says so to the person answering.

Voice agents are a named capability billed by the minute. That means a prospect picks up a call and speaks with software, and two passes located no disclosure requirement, no statement that the agent identifies itself as automated, no configuration governing it and no consent mechanism.

The same silence covers the other channels. Agents send email, text messages, web push notifications and messaging platform contacts, all generated and dispatched without a person composing them, and nothing indicates authorship to any recipient.

The unsubscribe allegation recorded on the compliance axis compounds this specifically. A person contacted by an autonomous agent, according to one detailed customer account, may have had no working mechanism to stop the contact, which removes the one control a recipient normally holds.

Identity itself is not falsified. Outreach carries the customer's own identity rather than a persona, and no rented accounts or fabricated senders appear anywhere.

The distinction that matters here is between who appears to be contacting someone, which is honest, and what is doing the contacting, which is not disclosed on any channel including the telephone.

Ask whether voice agents identify themselves as automated at the start of a call.

A published interface, a model context protocol server, and an open source framework with a real developer following behind it.

The protocol server is named explicitly in independent coverage alongside the interface as the two ways a customer feeds their own tools from the platform's context graph. Shipping one puts this among a small group in this index, and pairing it with a documented interface means the platform is designed to be built on rather than only used.

The open source origin is the deeper point. The underlying agent framework carries roughly sixteen thousand stars and two thousand forks on the public code host, which is a genuine developer community rather than a marketing claim, and it means the architecture beneath the commercial product is inspectable in a way almost nothing else here is.

Within the platform sit more than twenty five applications and agents, a studio for assembling go to market plays, a workspace, a plugin for the professional network and native integrations.

Breadth is corroborated externally: the product appears across nineteen categories on the principal review platform, which is unusual and reflects the sales, marketing and support modules operating as separate products.

Ask what the protocol server exposes and what the interface rate limits are.

Neither vendor surface could be read and nothing about the infrastructure was recoverable.

The marketing domain returns page metadata and a tag manager reference with no body content, indicating a client rendered application that a fetch cannot reach. The product domain disallows automated access outright. So no hosting provider, region, data centre, residency commitment, tenancy or isolation model, encryption statement, backup position, continuity plan or status page could be located from any source.

That is a retrieval limitation rather than a confirmed absence, and the grade should be read that way. A buyer with a browser will see material this record could not.

One indirect signal exists and cuts against the silence. A third party listing claims compliance with health information privacy rules, which cannot be achieved without defined infrastructure controls, documented processing locations and contractual undertakings. If that claim is accurate the underlying position is likely much better than this grade implies, and a buyer should ask for the documentation rather than assume from this record.

The holdings make it worth resolving: a database in the hundreds of millions, conversation transcripts, call recordings implied by voice agents, and a graph processing over a billion data points daily.

Ask where data is processed and whether any regional option exists.

Three credentials claimed, all by a third party, none verifiable in this pass.

An independent listing states the platform meets the European data protection regulation, a service organisation control examination and health information privacy rules, and describes it as suitable for regulated industries on that basis. If accurate that combination is stronger than most vendors of this size hold, and health information handling in particular is a demanding standard that would place this well above the norm for this index.

The difficulty is verification. Neither vendor surface could be read, so no trust page, security page, certificate, auditor name, scope statement, report request process, subprocessor list or vulnerability disclosure route is recorded here. The claim reaches this record entirely second hand.

One circumstantial support exists: a named customer operates in healthcare, and a vendor serving that sector would ordinarily have been required to evidence the relevant compliance during procurement.

Against that sits the review evidence, where a customer describes security and legal risk arising from missing functionality, though that concerns product capability rather than the vendor's own controls.

A buyer should treat this axis as unresolved rather than as either strong or weak, and settle it directly.

Ask for the examination report, its scope and its date.

Two different pricing models run on two different domains, and a buyer cannot tell which applies.

One surface carries team based flat rates at a hundred and fifty, three hundred and fifty and a thousand dollars monthly. The other carries a per seat model at forty nine dollars per user. Independent analysis notes both exist and describes the per seat arrangement as applying when user level billing is managed, which is not a distinction a prospective buyer would derive unaided.

What is disclosed well is the burn rate. Credits are consumed per action at published rates: one credit for an address enrichment, ten for a telephone enrichment, fifteen per minute for a voice agent. Publishing per action consumption lets a buyer model what a campaign actually costs, and few vendors metering on credits do it.

Billing mechanics are documented in unusual detail: seats charged in full on addition with no proration, automatic seat purchase when an administrator invites someone with no free seat remaining, seat reduction effective only next cycle and never below active membership, credit packs charged immediately and expiring at cycle end, and a five day payment retry before access is paused.

Two things undercut it. Credits expire monthly with no rollover, so unused capacity is forfeited. And automatic seat purchase means headcount changes generate charges without an explicit decision.

A permanent free tier at a hundred credits monthly exists.

Ask which pricing model applies and whether credits ever roll over.

Two programmatic routes out, and one customer account describing a contractual exit that was anything but clean.

The technical position is reasonable. A published interface and a model context protocol server both provide extraction paths, and the record system is the central store, so data is consolidated rather than scattered across modules. A customer with development capacity has somewhere to start.

Two passes located no export mechanism described in product terms, no format, no scope statement, no deletion commitment and no retention position after cancellation.

The contractual side is where the concern sits. One detailed customer review describes an annual agreement, an inability to use the platform throughout the contract period, and having to escalate to the founders to terminate and recover a refund. That is a single account among four hundred and forty one reviews and should be weighted accordingly, but it describes the exit process rather than the product, which is what this axis measures.

Credits expiring monthly with no rollover means unused capacity is forfeited continuously rather than at exit, so nothing accumulates that would be lost on departure.

Ask what the standard contract term is and what the cancellation terms are.

Three independent sources agree on the same three absences, and competitors attack on exactly this.

No warm up engine. No domain rotation. No address control. Multiple analyses state that teams running cold outreach through the agents get no deliverability infrastructure at all, and one frames it as the point where the platform's limitations show once volume or reply quality start to matter.

That gap sits awkwardly against what the platform does. Autonomous agents send email, text messages, web push and messaging platform contacts continuously, which is precisely the volume profile that requires warm up, rotation and reputation management to survive.

The unsubscribe allegation recorded on the compliance axis belongs here too, because unsubscribe handling is a deliverability mechanism as well as a legal one. Complaint rates are among the strongest signals filtering systems weight, and a recipient with no working stop mechanism complains instead.

One customer review describes months without being able to send a single campaign, attributing it to defects, data problems and misconfiguration.

What could not be established, because neither vendor surface was readable, is whether any authentication guidance, bounce handling or placement testing exists undocumented in third party coverage.

Ask what deliverability infrastructure exists for agent driven sending.

Nineteen categories on the principal review platform, which is the broadest footprint in this batch.

That number reflects genuine product breadth rather than category gaming. Sales, marketing and customer support each operate as substantial modules, and the platform is reviewed separately across all three, which is why it appears under two profiles and nineteen categories.

The named customer set spans developer tooling, communications infrastructure, collaboration software, healthcare services and security, which is a wider industry spread than most records here and suggests the consolidation pitch lands across sectors rather than in one vertical.

The commercial range runs from a permanent free tier at a hundred credits monthly through to a thousand dollars monthly on the flat rate model, so the ladder covers an individual experimenting through to a business team.

The tension worth recording is between positioning and assessment. The vendor presents an artificial intelligence native record system unifying sales, marketing and service for organisations of any size, while independent analysis places it as best suited to smaller commerce and direct to consumer teams wanting those functions bundled. Those are different buyers.

No geography, no vertical focus and no company size band was located.

Ask which segment the customer base actually concentrates in.