Methodology

Agentic Index tracks and compares 969 vendors offering AI agent platforms, autonomous workflow automation tools, AI coding agents, GTM and revenue agents, browser/computer-use agents, customer support agents, and enterprise operations agents. We define an "agentic AI product" as software where an AI model takes multi-step actions, uses tools, and operates with some degree of autonomy toward a goal, not just a chatbot or single-turn inference product. Products must be shipping or in documented public beta to be included. Research previews and unpublished internal tools are excluded unless they have a public waitlist or published documentation. Canonical inclusion requires three criteria: (1) publicly documented agentic functionality: evidence that the product takes multi-step autonomous actions, uses tools, or orchestrates workflows across systems; (2) a practical evaluation path: a self-serve trial, free tier, open-source install, or documented API/SDK that allows a buyer or developer to evaluate the product without a mandatory sales call; and (3) enough source specificity to avoid vague marketing: at least one primary source (official docs, pricing page, trust center, or official changelog) with concrete feature or deployment evidence. Products with only general-purpose AI assistant positioning, unclear category fit, weak governance documentation, or no practical public evaluation path are tracked on an internal watchlist and not included in the public directory.

Every grade rests on the vendor's own published materials. In descending order of weight: • Legal pages: terms, privacy policy, DPA, subprocessor list • Help centre, documentation, developer portal and changelog • Product, pricing, security and trust pages • The vendor's own blog and announcements • Marketplace listings the vendor authored and submitted, graded for capability and named as the channel; pricing on a listing is channel specific and never stands for the direct offering • Public repositories the vendor maintains First party is decided by authorship and legal responsibility, not by where a page is hosted. Third party material is not grading evidence: analyst reports, reviews, listicles, press coverage, aggregator profiles and competitor comparison pages never carry a grade, and where a cell's only support is such a source the cell is recorded as Unknown rather than as a grade. Company facts such as funding, headquarters and founding date may be taken from press announcements quoting the vendor, and sit in descriptive fields rather than in any grade. We do not rely on vendor-provided data submissions. We do not accept payment for inclusion or favorable positioning. Vendors are not contacted for data verification at this stage, though that may change.

The feature taxonomy is organized into six categories derived from the buyer-facing agentic AI capability rubric: • Action & orchestration: Integrations & Tool Calling, Workflow Orchestration, and Triggers & Channel Coverage. • Knowledge & context: Knowledge Grounding & RAG, and Memory & State Persistence. • Control & trust: Human Oversight & Guardrails, Security Identity & Governance, Observability & Auditability, and Deployment & Data Residency. • Solution readiness: Prebuilt Agents, Templates & Packs. • Platform extensibility: Model Flexibility & Routing, APIs/SDKs & MCP Extensibility, and Testing, Debugging & Optimization. • Specialist automation: Browser & Computer Use. The taxonomy covers 14 canonical features ranked by buyer importance (importanceRank 1 to 14). Each feature carries equal weight (1/14, roughly 7.1%) in the coverage score calculation. One fact does work on one axis only: sovereign or on premises delivery is credited on deployment and never also on security, a knowledge graph credited on grounding is not also credited on memory, and audit logs credited on observability do not also clear security. Feature definitions are fixed at taxonomy publication; when a definition is refined, the revision is dated in the rubric revisions section below and affected vendor records are re-verified and dated.

Each vendor receives a feature coverage score calculated from its 14 vendor-feature records. Scores use the following point values: • Full / Explicit (F) = 1.0 point: confirmed from primary sources. • Partial (P) = 0.5 points: limited or restricted capability documented. • No / Not documented (N) = 0.0 points: confirmed absent from public materials. • Unknown / Unspecified (U) = 0.0 points: not yet verified from primary sources. The coverage score is the sum of point values divided by 14 (maximum). This ranking reflects documented public-source coverage, not a subjective quality judgement. A high Unknown count depresses a vendor's score until research is completed. It does not indicate poor capability.

Each vendor-feature relationship is assigned one of four support levels: • Full / Explicit: We have confirmed from primary sources that the vendor supports this feature explicitly. • Partial: The vendor supports a limited or restricted version of the capability; a note describes the limitation. • No / Not documented: We have confirmed from primary sources that this capability is absent or not publicly documented as first-class. • Unknown / Unspecified: We have not yet verified this vendor-feature combination from primary sources, or the only material found sits on sources the methodology excludes. This is the default state and should not be interpreted as a definitive absence of capability. A capability purchasable at any tier, including a quoted enterprise tier, is graded as documented with the tier named. A capability reachable only through closed early access, a waitlist or an application is Partial with the gate named. Roadmap statements, coming soon language and future tense announcements are not evidence of shipped capability. Neither a vendor's claim nor a vendor's disclaimer decides a grade; the documented mechanism does. Full and No ratings require a source URL and a confidence level.

Individual vendor-feature records carry a confidence score. Confidence describes how well the vendor's own evidence supports the grade; it never substitutes for first party evidence, which is why third party sources appear in no tier. • High: The vendor's documentation, legal pages, trust centre or a product page names the mechanism. A named single region or a named single provider is a high confidence None, because the absence is visible rather than inferred. • Medium: A vendor blog post or announcement, or a gated developer or trust surface whose existence and scope are documented but whose detail cannot be read. A Full resting on a gated surface is held at medium. • Low: A first party statement with no mechanism described, or evidence that could not be re-retrieved at the last verification. Where the vendor's own surfaces contradict each other, the contradiction is recorded in the evidence note and moves confidence rather than the grade. Vendor-level records also carry a confidence level reflecting overall data completeness. A vendor with many Unknown feature cells will have a lower overall confidence rating.

Each vendor record and vendor-feature record carries a lastVerifiedAt date. Data older than 6 months should be considered potentially stale in fast-moving categories. Records that have not yet been formally reviewed may lack a lastVerifiedAt date. The absence of a verification date means the data should be treated as a research starting point, not a definitive assessment.

Since 30 August 2026 every record has been under a lane by lane accuracy review against the bars published on the feature pages. Each reviewed record carries a dated review line in its notes stating what was confirmed, corrected or left unresolved, and every corrected cell names the page it now rests on. Where a bar was found stated differently across documents, the published feature page was corrected and the date recorded here. • 1 September 2026: Integrations, Memory, Testing and Browser bars restated to match the published feature pages, replacing working paraphrases that had drifted stricter or looser. • 5 September 2026: Prebuilt Agents no longer requires a browsable library for Full; the test is whether the packaged assets are separately adoptable whole products. Security requires both a documented access surface and a compliance posture for Full, either half alone being Partial. • 6 September 2026: Deployment counts a one sentence residency option with no named region as Partial and a single disclosed region with no choice as None. Extensibility credits a gated developer portal at Full where the API is named with its scope and access path, and grades on the direction of the call. • 7 September 2026: Testing counts a vendor operated evaluation harness at Full where the artifact under test is the buyer's own deployed agent. Model Flexibility page restated to say the axis measures who chooses the model, not how many are involved. Grades made before a revision are re-verified as their lane is reviewed and carry the date of that re-verification, so a record's verification date states which version of the bar it was graded under.

The Agentic Index is not: • A subjective recommendation engine: rankings reflect documented public-source coverage, not editorial opinion. • A paid directory: vendors cannot pay for inclusion, placement, or score inflation. • A vendor-verified database: vendors do not approve their own entries. • A real-time monitor: data is updated periodically, not continuously. The index is designed as a structured research resource for buyers, analysts, and operators who need to evaluate the agentic AI landscape based on documented capabilities rather than marketing claims.