AgentenCode does not try to make the largest possible number of claims about an AI agent. It tries to make it clear which claims are supported, how current they are, what scope they apply to and where editorial interpretation begins.
AI-agent products change quickly. Capabilities are added, pricing and plans are reorganized, models are replaced, integrations are renamed and enterprise controls can appear only in specific regions or tiers. AgentenCode therefore treats a profile as a traceable research state at a particular point in time — not as a permanent verdict.
A precise gap is better than a falsely complete statement. If information cannot be supported reliably, AgentenCode does not guess. Unknown remains Unknown.
1. Four states for traceable claims
Facts, provider statements and editorial interpretation must not collapse into one label. AgentenCode therefore uses distinct claim states that apply to individual fields or statements rather than to an entire product.
A concrete statement is supported by a current, traceable primary source. “Verified” does not automatically mean AgentenCode technically tested or audited the product.
A feature or property is described in official documentation, while scope, prerequisites or limitations may still be incomplete.
No sufficiently specific current primary source is available, or public information is ambiguous or contradictory. This explicitly does not mean “No”.
A summary, category or practical interpretation based on documented facts. It is not presented as a provider claim.
2. Sources first — with a clear hierarchy
For changing product properties, primary sources have priority. These include official product pages, technical documentation, help centers, developer documentation, release notes, security and privacy documentation, legal terms and official provider announcements.
AgentenCode tries to verify a claim as close as possible to the original source. Secondary sources can add context, point to a new change or help explain a historical development, but they do not silently replace a suitable primary source for current product capabilities.
Source hierarchy
- Official technical documentation and product specifications
- Official product, security, privacy, legal and support documentation
- Official release notes, roadmaps and provider announcements
- Reliable secondary sources for context or discovery
- Editorial interpretation based on documented facts
3. How an Agent Profile is created
A detailed AgentenCode profile follows a repeatable workflow. The goal is not to force every product into an identical template, but to answer comparable core questions consistently.
- Resolve identity: provider, product name, product type and current lifecycle status are identified unambiguously.
- Collect primary sources: relevant official pages are mapped to the fields they can actually support.
- Capture core attributes: capabilities, autonomy, models, tools, memory, integrations, protocols, platforms, deployment, privacy, security, enterprise controls, pricing and source access are reviewed.
- Mark uncertainty: missing or contradictory information is not replaced by assumptions.
- Add editorial context: practical use cases, strengths and limitations are derived from documented properties and clearly treated as interpretation.
- Set a verification date: every profile carries a traceable date for its last substantive review.
- Publish sources: important statements should remain traceable to the official evidence behind them.
4. What AgentenCode examines systematically
Agentic products vary widely, but profiles use a shared conceptual model so readers do not have to learn a completely different logic on every page.
Product & capabilities
- Provider and product type
- Target audiences and use cases
- Autonomy and approval steps
- Models and model selection
- Tools and external actions
- Memory and context
- Integrations and protocols
- Outputs and source access
Operations & control
- Platforms and deployment
- Open-source status and licensing
- Enterprise controls
- Privacy and data use
- Security capabilities
- Pricing and access
- Strengths and limitations
- Verification date and change history
5. Product facts and editorial interpretation stay separate
One of AgentenCode’s central jobs is to keep verifiable product properties distinct from useful but interpretive conclusions.
Example: documentable statement
“The agent can access specific external tools.” That statement should be traceable to provider documentation that explicitly supports the capability and its scope.
Example: editorial interpretation
“The agent is particularly suitable for complex research workflows.” That may be a reasonable conclusion from documented capabilities and limitations, but it is not the same kind of fact as a provider-documented feature.
AgentenCode may explain and classify a product. The important requirement is that readers can see when the platform is documenting evidence and when it is interpreting it.
6. Freshness: a verification date is not an expiry guarantee
AI-agent products are among the fastest-changing software categories. Detailed profiles therefore show a last substantive verification date. That date means the important publicly available information was reviewed at that point. It does not mean every capability remained unchanged afterward.
Pricing, quotas, model availability, staged rollouts, regional availability and preview features are especially volatile. For time-sensitive decisions, the linked provider source remains authoritative.
AgentenCode is designed to shorten the path from a claim to its supporting source, not to hide the source behind a score.
7. Structured states and versioned history
AgentenCode does not keep product knowledge only as prose. Alongside visible profiles it maintains structured records that connect agents, providers, sources, normalized claims and change events with stable identifiers. We refer to this evidence graph as AgentenGraph.
The value is not that public provider facts become proprietary. The value comes from normalization, time context and traceable state transitions. When a structured value changes, the earlier state should not simply disappear. Where the evidence model supports it, the old state is closed, the new state is recorded and the transition becomes a history event.
Field-level evidence is never pretended
A source attached to a profile is not automatically treated as proof for every field in that profile. A claim becomes field-level evidence only when a concrete source actually supports that field for the relevant scope.
This creates a durable baseline for later comparisons across capabilities, autonomy, models, tools, MCP/A2A, APIs, integrations, hosting, privacy, data use, enterprise controls, pricing, regions and certifications.
8. System level: agent, framework, component, system and institution
Modern agentic products are increasingly composed of multiple technical and organizational layers. AgentenCode therefore distinguishes between an agent, an agent framework or harness, an agent component, an agent system and an agent institution.
Relations are not guessed
A statement such as “Agent A delegates to Agent B”, “Agent X uses model Y” or “this platform control governs this agent” is stored as a documented relation only when a specific primary source supports that relation for the relevant scope. Platform features are not automatically inherited by every agent, and transitive relations are not inferred by default.
This system model prepares evidence fields for composition, multi-agent orchestration, delegation, handoffs, agent-to-agent communication, shared memory, tool permissions, human approval, rollback and cross-agent tracing. Fields are populated only when primary evidence is available.
9. AgentenWache: monitoring primary sources for change
Registered primary sources are monitored technically for meaningful changes. The internal monitoring layer, AgentenWache, can compare reachability, redirects and normalized content fingerprints to identify pages that deserve editorial review.
A detected delta is not yet a new fact
Monitoring never publishes a product claim automatically. A technical difference first enters review. Only when the current primary source supports a substantive change is the profile, evidence field or history layer updated.
Automated monitoring has limits: providers may block automated requests, dynamic pages can generate irrelevant changes and some documentation is only accessible after authentication. Those limits are not bypassed, and a monitoring signal is never treated as equivalent to a confirmed product change.
News publishes only confirmed, relevant changes
News is not an automated AI-news feed. A change becomes an article only after editorial verification, when it matters to the reference layer and can be linked to an affected profile and primary source. Rumors, funding-only stories and generic AI news without a concrete agent change are excluded.
10. Comparison without an artificial overall winner
A research agent, a coding agent and an enterprise orchestration platform solve different problems. AgentenCode therefore avoids pretending that one universal score can rank them fairly.
Comparisons focus on concrete dimensions such as autonomy, tool access, deployment, source access, protocols, enterprise controls, privacy options and documented governance. When an ordering is shown in a tool such as Agent Fit, it is tied to the criteria selected by the user rather than a hidden editorial “best agent” score.
11. No paid rankings
Editorial independence
A payment cannot buy a better editorial position.
AgentenCode does not sell better rankings or more positive evidence states. If sponsorships, partnerships or commercial provider formats are introduced, they must remain clearly labelled and separated from the evidence-backed reference layer.
Providers can submit information or corrections. Whether and how those changes reach a profile depends on evidence and methodology — not on a commercial relationship.
12. What AgentenCode explicitly does not claim
AgentenCode is an evidence-backed editorial reference, not a certification body, legal adviser, penetration-testing lab or independent benchmark authority. Public documentation can be incomplete, and a provider can change a product between verification cycles.
- A documented security feature is not a security certification by AgentenCode.
- A control mapped to the EU AI Act does not prove legal compliance.
- Data residency does not by itself prove GDPR compliance.
- A documented model or benchmark statement is not an independent performance test.
- Unknown does not mean absent.
- A profile is not a substitute for contractual, security or legal due diligence.
13. Changes, corrections and new evidence
When new evidence appears, AgentenCode updates the relevant field, source reference or profile text and, where appropriate, records a history event. Corrections should preserve traceability: the goal is not merely to display today’s answer, but to make it possible to understand why a material statement changed.
Provider corrections are welcome when they include a primary source or sufficiently specific official documentation. If the source is ambiguous, the conservative outcome is to keep the field Unknown rather than infer a stronger statement.
14. Machine readability, SEO and generative search
Human-readable pages, structured data and public JSON are intended to describe the same underlying state. AgentenCode publishes canonical URLs, reciprocal hreflang annotations, schema.org JSON-LD, sitemaps, llms.txt and selected public datasets so search engines, developers and AI systems can resolve evidence consistently.
GEO optimization does not mean inventing answer-shaped claims. The preferred pattern is a concise direct answer followed by scope, evidence, verification date and primary sources. Machine-readable output must follow the same Unknown and source rules as the visible site.
15. How AgentenTrust and History work together
AgentenTrust normalizes governance evidence into 36 controls across privacy, security, human control and auditability. The current state answers “what is documented now?” The history layer answers “what verified change do we have stored over time?”
A trust control can be documented true, documented false, carry another documented value, or remain unknown. Field-level history stores a transition only when a previous and a new state can be supported after the tracking baseline. Earlier states are not retroactively invented.
16. Daily revision, review and publication
The daily workflow is one shared evidence process for German and English. Sources and changelogs are checked once. Confirmed facts update the shared data layer once. The German and English pages then render the same evidence in different editorial language.
This avoids two dangerous failure modes: maintaining competing databases for each language, or translating stale prose without updating the underlying evidence. The factual record remains single-source; language presentation is derived from it.
17. Public reference and private review remain separate
Monitoring signals, raw review notes and unverified candidate changes are not published as facts. The public layer contains sanitized, reviewed evidence. This boundary prevents internal monitoring noise from leaking into profiles, search results or machine-readable datasets as if it were verified product knowledge.
Read-only API, Fact Explorer and MCP access
Static JSON API
Selected datasets expose profiles, evidence, governance, history and relations in machine-readable form without creating a second factual store.
Queryable facts
Ask, Check, Compare, Fit and History Query read from the same public evidence layer used by the profile pages.
MCP
Machine access should preserve the same scope, source and Unknown semantics as the human interface. A structured interface does not make uncertain data certain.
Frequently asked questions about the methodology
What does “Verified” mean?
It means a concrete statement was checked against a current, traceable primary source. It does not automatically mean AgentenCode independently tested or audited the product.
What does “Not confirmed” or “Unknown” mean?
It means sufficiently specific current evidence is missing, ambiguous or contradictory. It does not automatically mean the capability does not exist.
Can a provider pay for a better ranking?
No. AgentenCode does not sell editorial ranking positions or positive evidence states.
How does AgentenCode distinguish an agent from an agent system?
The system taxonomy separates the acting agent from frameworks, components, composed systems and broader institutional arrangements. Relationships are stored only when explicitly supported.
How current are Agent Profiles?
Profiles show a substantive verification date, but products can change after that date. For time-sensitive details, the linked primary source remains authoritative.
Why Ask AgentenCode is not a normal chatbot
Ask AgentenCode does not try to answer from general model memory. It evaluates questions against the published evidence corpus and returns matches, unknowns, scope and primary-source context. If the structured data cannot support an answer, the correct response is to show that limitation rather than fill the gap with a plausible guess.