Go back to Market Research
AGENTS.inc, operated by Berlin-based AGENTS HQ GmbH, supplies software agents for company discovery, market research, promotional content, and sales qualification. Its AGENTS HQ platform coordinates specialized agents, connects them to enterprise applications and external information, and produces reports, editable layouts, or customer interactions. Its supply chain application is principally supplier discovery and monitoring. (1, 2, 9, 18)
Overall supply chain score: 3.2/10
| Dimension | Score |
|---|---|
| Supply chain depth | 2.8/10 |
| Decision and optimization substance | 2.6/10 |
| Product and architecture integrity | 4.2/10 |
| Technical transparency | 3.2/10 |
| Vendor seriousness | 3.0/10 |
AGENTS.inc automates information gathering and document production, with some workflows also updating business systems. A September 2026 supplier-search trial provides concrete evidence of reduced research time. The public product does not calculate inventory policies, replenishment quantities, or production decisions under uncertainty. Its low supply chain score reflects that limited operational scope as well as weaknesses in its technical and economic claims.
The most consequential weakness is the gap between promised autonomy and required verification. The accuracy page claims zero hallucinations in published executive reports and perfect response accuracy in an insurance deployment. The terms acknowledge factual errors and put responsibility for checking every output on the customer before publication or public use. That continuing review obligation belongs in the labor calculation behind the claim of an autonomous workforce. (3, 6)
The vendor describes coordination, model selection, validation, containerization, and deployment choices. It also includes integration, retuning, and maintenance in the base fee. The weaknesses lie in the claimed guarantees: source attribution does not ensure correct reasoning, and changing models can alter application behavior. Parallel execution reduces elapsed time while adding computing and validation costs that the public scalability account leaves unquantified. (2, 5, 8)
AGENTS.inc versus Lokad
AGENTS.inc helps a procurement team discover candidate suppliers and collect evidence about them. Its Company Agent accepts criteria and returns structured company lists. Buyers then compare candidates and select suppliers. This can shorten a costly preliminary step, especially when the purchasing team searches unfamiliar markets. (9)
Lokad addresses the subsequent economic decisions: quantities, timing, sourcing allocations, inventory positioning, and other actions whose value depends on uncertain demand and supply. Supplier information can inform those decisions, but it does not resolve them. A manufacturer may discover an additional certified supplier and still need to decide whether a lower price compensates for a longer, less reliable lead time.
That distinction is substantial. An agent can accurately reproduce a supplier’s advertised certification while missing a qualification that matters for the customer’s product. It can correctly identify a nominal delivery time without estimating the distribution of actual deliveries. It can assemble a convincing shortlist without determining which allocation of purchases produces the best expected economic result. AGENTS.inc’s public sourcing account ends at the shortlist and the buyer’s selection.
The operating comparison therefore concerns complementary activities rather than equivalent planning engines. AGENTS.inc’s task-based service can replace research and production steps around a decision. Lokad’s design perspective, developed in Introduction to Supply Chain, aims to maintain the economic logic that generates decisions repeatedly. Both approaches require integration and specialist work; the relevant measure is the full recurring labor and cost associated with the output actually delivered.
Supplier discovery: the strongest supply chain evidence
EU-LAC’s September 24, 2026 Machu Picchu Foods sourcing trial reports 3,136 companies screened and 81 relevant suppliers across 16 searches. Search time fell from an approximately eight-hour baseline to at most one hour, often fifteen to twenty minutes. Buyers still selected suppliers. (9)
Predefined targets included twenty queries, more than 80% less active research time, relevance above 4/5, and actionable shortlists in over 85% of cases. The results instead cover sixteen searches, each finding at least one relevant supplier. Final average relevance and achievement of the targeted three to five strong leads per actionable shortlist are not reported.
The evidence supports faster supplier discovery in a bounded trial. It does not establish a reduction in purchase prices, better delivery reliability, successful supplier qualification, or lower total procurement cost. Wider rollout and integration into daily procurement are described as next steps. The distinction between elapsed search time, active buyer time, and total qualification effort matters: a fast shortlist can still require substantial commercial and technical investigation.
The case also illustrates why search depth is not an economic objective. Screening more companies increases potential coverage, but can add duplicate candidates, weak matches, and review work. The right stopping rule depends on the expected value of another credible option relative to the cost of finding and qualifying it. AGENTS.inc exposes search-depth choices; it does not publish a decision rule that balances those costs.
What the agents execute
The platform has three described layers: specialized agents, a coordination hub for routing and retries, and a validation layer. Agents can pass research to a reporting agent and then to presentation generation without a person relaying each step. Model choice occurs inside an agent, with commercial and local model options. (2)
Some outputs extend beyond summaries. The integrations page describes reading promotional records from SAP PMR, writing qualified leads and conversation history to Salesforce, booking meetings through Office 365, and generating editable InDesign files. These are identifiable application actions. They are not evidence of economic inventory optimization, but they do show that the product’s automation is broader than a conversational search box. (7)
The promotional workflow is particularly concrete. One agent generates copy and layout from product data; a separate check compares printed prices with their source records. The vendor reports 80% faster production for a national retail deployment. The customer is not named on that case page, and the measurement period, denominator, and residual designer effort are not disclosed. An editable document is the output; a campaign’s incremental margin is a different outcome. (10)
Market intelligence and innovation scouting follow a recurring pattern: define criteria, identify companies, monitor information, assess fit, and generate a report or deck. This can remove repetitive collection and formatting work. It also leaves the choice of criteria, interpretation of incomplete evidence, and economic decision with the customer. Ten parallel market reports are not ten validated forecasts of demand. (12, 13)
Source attribution is not a proof of correctness
AGENTS.inc’s accuracy explanation contains one strong, bounded mechanism: checking a generated price against the corresponding SAP record. With a defined product identifier, currency, date, tax treatment, and promotion scope, a deterministic comparison can reject a mismatched number.
The same explanation then extends to every factual claim in open-ended reports. It describes independent validation, provenance, escalation, and zero hallucinations across published executive reports. The public material supplies no evaluation set, error taxonomy, sample size, rejection rate, or method for establishing that a source actually supports the generated claim. (3)
These omissions concern the mechanism, not just documentation polish. A real source may be outdated, refer to a different legal entity, contain a vendor’s unsupported claim, or describe a target rather than an achieved result. A sentence can cite that source and still be wrong. A separate model reviewing a statement can share the first model’s errors; independence of execution does not establish independence of failure.
Nor does refusing unsupported statements establish completeness. An agent that omits the difficult evidence can produce an apparently flawless report while failing its purpose. Evaluation must include what the system missed, the cases escalated to people, and the work discarded by validation. Reporting error rates only on accepted outputs can conceal these costs.
The terms provide a more restrained account. Section 5 says the validation layer reduces hallucinations, acknowledges factual and formatting errors, and requires customers to fact-check and approve all outputs before publication or public-facing use. This is a direct inconsistency with the stronger proposition that the architecture makes autonomous output safe to publish. The claimed verification benefit must be measured after the customer’s remaining review. (6)
Customer scope and the perfect-accuracy claim
The insurance case identifies Fink & Wagner AXA and Salesurance as the client. Alongside that identification, it displays AXA’s group-wide scale of 94 million clients across fifty countries and describes thousands of prospects handled daily. A deployment at an agency should not inherit the scale of the entire insurance group without evidence of a wider rollout. (11)
The same page claims 100% response accuracy and no risk of incorrect information. It does not disclose the assessed conversations, how correctness was judged, which questions were refused, or whether that percentage covers live production traffic. A tightly restricted answer set can be checked much more rigorously than unrestricted insurance advice; the page does not define that boundary.
The appropriate test is a representative, independently reviewed sample of conversations, including difficult questions, escalation, and changes in the questions customers ask.
Technology history and follow-through
The agent theme predates the current LLM wave. That history establishes continuity of product ambition, but it also requires distinguishing earlier information-retrieval automation from present claims of an autonomous workforce.
| Period | Public record | Assessment |
|---|---|---|
| 2013–2016: collaboration software | Sebastian Denef’s 2018 account describes a visual collaboration product, difficult commercialization, and a shift toward agents during the 2016 SAP/Capgemini program. | A founder’s account of changing product direction, rather than evidence that today’s agent capabilities existed then. |
| 2017–2018: agent framework | The same account reports a running framework, an Innogy investment, and hackathon prototypes for scientific work, meetings, and ICO analysis. | Concrete development activity. Prototype demonstrations do not establish durable production automation. |
| 2019: research assistants | An interview published by the company’s technology-park host describes news, science, company, and intellectual-property agents, with a shared visual board. | The basic research workflow is recognizable in the current product. The assertion of replacing weeks of research with minutes lacked a controlled evaluation. |
| 2021-era platform launch | Scaleway’s customer account describes migration from AWS, containerized agents, and a planned marketplace where agents would seek contracts. | Attributable infrastructure work and a broader platform ambition. The current offer centers on vendor-built, maintained agents; delivery of the proposed marketplace is not explained. |
| 2024: infrastructure work | An engineer’s public CV attributes Kubernetes migration, ArgoCD, and Terraform work to AGENTS.inc. | Specific engineering evidence beyond successive marketing labels. |
| 2026: managed workflows and sourcing trial | Current pages emphasize managed marketing and sales automation; EU-LAC documents a procurement-search proof of concept. | A narrower sourcing result can be assessed. The universal reliability claims still lack comparable evaluation detail. |
Pre-2023 references to AI agents should not be read as evidence of mature general reasoning. Searching databases, classifying material, or assembling summaries can be implemented with narrower methods. The 2019 technical record includes Python and scikit-learn work, consistent with that more limited scope. The relevant question is what was automated and validated at the time, not how closely its label resembles a current LLM agent.
The current claim that changing the model behind an agent changes nothing else in the pipeline is too strong. An interface can remain stable while extraction accuracy, refusal behavior, tool selection, latency, and cost change. Model substitution requires regression testing against the application’s real tasks. AGENTS.inc does not publish that test process or its results. (2)
Bill of software materials
Current public pages disclose deployment and model choices. Earlier infrastructure material and an engineer’s account provide more specific components.
| Layer | Attributable materials and date | Operating implication |
|---|---|---|
| Research-agent implementation | Python and scikit-learn in an engineer’s account of work during June–December 2019. | The earlier agents included conventional data preparation and machine-learning components. |
| Historical data platform | Scaleway’s launch-era architecture diagram shows Kafka, PostgreSQL, object storage, Elasticsearch, and MongoDB, with application and agent services across two Paris zones. | Several persistent data systems require synchronization, backup, access control, and coordinated changes. |
| Deployment automation | Kubernetes, ArgoCD, and Terraform attributed to July–November 2024 work; current security material independently confirms Kubernetes. | Orchestration and infrastructure management are maintained software responsibilities. |
| Model execution | Current platform material identifies Gemini, Claude, GPT, Mistral, and local open-weight options. | Behavior, data routing, availability, and inference cost depend partly on external model services or customer-hosted alternatives. |
| Business-system connections | SAP PMR, Salesforce, Office 365, InDesign, and WhatsApp interfaces. | The service must maintain the meaning of product records, permissions, document formats, and workflow actions as those systems change. |
The launch-era diagram shows substantial infrastructure behind the agents. Its multiple data stores require consistency checks, recovery procedures, and coordinated upgrades. AGENTS.inc does not publish which components remain, their current responsibilities, or the operating cost per completed, checked task.
The deployment page says that ten sector analyses take no longer than one, and eight markets no longer than one. Parallel execution can reduce elapsed time when enough resources are available. It does not remove the additional inference, retrieval, storage, validation, and coordination work. Throughput should be reported with the workload and resource cost; elapsed time alone is an incomplete scalability claim. (8)
Security requires more than residency
The security page describes dedicated workspaces, containerized deployment, customer-hosted options, and contractual restrictions on model-provider retention. These address parts of the data-handling problem. It does not explain whether tenant isolation is enforced through separate clusters, namespaces, service identities, databases, or application rules. (4)
Residency is also described imprecisely. The security page names Scaleway in France and STACKIT while promising strictly German processing for the relevant European deployments. The privacy policy separately lists cloud and model subprocessors, including Azure in the EU or USA. A customer needs the actual route of its records, prompts, logs, and model requests; a provider’s nationality cannot establish that route. (4, 19)
The application attack surface extends beyond hosting. Agents retrieve external material and can interact with mail, calendars, and CRM records. A hostile source can carry instructions aimed at redirecting an agent. The public documentation does not explain how untrusted retrieved text is separated from executable instructions, how credentials are kept outside model-visible content, or how action permissions are limited per task.
Checking factual claims after generation does not solve those authorization problems. An agent can quote a document correctly while sending it to an unauthorized recipient. Containerization can constrain a process, but its effectiveness depends on network permissions, mounted secrets, execution privileges, and the scope of business credentials. These boundaries deserve an engineering description before unattended access is treated as established security.
Configuration maintenance, productivity, and accountability
AGENTS.inc explicitly prices continuing work. The base fee includes onboarding, integration, validation-rule configuration, monitoring, maintenance, and retuning when source data or cases change. A usage charge then applies to agent work. This is a more complete operating proposition than a license that silently transfers all configuration upkeep to the buyer. Actual rates and the definition of a billable task are determined during scoping. (5)
The terms refer to Agent Tokens as the consumption unit. The practical cost therefore depends on how a business task maps to searches, model calls, retries, validation, and escalations. A per-task commercial description does not by itself guarantee predictable cost per accepted deliverable. Buyers need the task boundary, retry treatment, and quality acceptance criteria to make that calculation. (6)
The customer still supplies objectives, source-system semantics, brand or qualification rules, and expert judgment. These are executable business assumptions even when written as instructions or configuration rather than code. A new supplier certificate, changed promotion rule, or altered insurance product can require adjustments and retesting. Included retuning places some responsibility with the vendor, but public documentation does not show version control, rollback, test coverage, or how changed rules are validated before autonomous operation.
The Machu Picchu trial measures a defined piece of this work and reports a time saving. That is more informative than counting agents or reports. A full productivity account would also include qualification of proposed suppliers, rejected searches, expert reviews, vendor tuning, and any work shifted to another department. The proof of concept should be extended on that basis before its speedup is treated as a department-wide result.
Accountability remains asymmetric. The pricing page says the base fee transfers operating risk to the vendor. The terms put the final responsibility for output review on the customer and exclude damages from unverified output; separate enterprise agreements may provide additional service commitments. The public promise and the default obligation describe different levels of autonomy. (5, 6)
For supplier discovery, a useful economic test would follow qualified candidates through actual sourcing outcomes: successful qualification, purchase terms, delivery reliability, and the value of alternatives that become available. For campaign production, it would include correction and approval work alongside production time. AGENTS.inc’s current evidence is strongest on completing information-processing tasks faster, and much weaker on the economic quality of the resulting business decisions.
Supply chain score
Supply chain depth
| Criterion | Score |
|---|---|
| Economic framing | 3/10 |
| Decision end-state | 2/10 |
| Conceptual sharpness on supply chain | 3/10 |
| Freedom from obsolete doctrinal centerpieces | 4/10 |
| Robustness against KPI theater | 2/10 |
Average: 2.8/10
Economic framing — 3/10. Research time and service cost receive attention, and supplier discovery can have commercial value. The sourcing method does not connect its recommendations to margins, working capital, purchase quantities, or expected disruption losses. A cheaper search is an input to procurement performance rather than a complete economic result. (5, 9)
Decision end-state — 2/10. The supply chain workflow returns candidates for a buyer to assess. CRM writes and campaign files are concrete outputs elsewhere in the product, but the public offer does not generate purchase allocations, replenishment orders, or production decisions. (7, 9)
Conceptual sharpness on supply chain — 3/10. Supplier criteria include technical fit, certification, and geography. The method does not explain how supplier reliability, uncertain future needs, or dependencies between sources affect selection. A supplier’s published description cannot substitute for that analysis. (9)
Freedom from obsolete doctrinal centerpieces — 4/10. The offer avoids forecast-consensus meetings and inventory classifications because it operates upstream of those processes. Its limitation is the continued treatment of the shortlist and human selection as the endpoint. Freedom from traditional planning rituals has not produced an alternative economic planning method. (9)
Robustness against KPI theater — 2/10. The supplier-search trial reports useful quantities and a baseline. Elsewhere, zero hallucinations, perfect response accuracy, and group-scale customer descriptions are presented without the definitions needed to assess them. The narrow trial cannot validate these broader claims. (3, 11)
Decision and optimization substance
| Criterion | Score |
|---|---|
| Probabilistic modeling depth | 1/10 |
| Distinctive optimization or ML substance | 3/10 |
| Real-world constraint handling | 2/10 |
| Decision automation and labor productivity | 5/10 |
| Resilience under real operational complexity | 2/10 |
Average: 2.6/10
Probabilistic modeling depth — 1/10. No demand or lead-time distribution, dependency model, or stochastic economic decision calculation is documented. The probabilistic behavior of an LLM, acknowledged in the terms, is a different matter from modeling uncertainty in future supply chain events. (6)
Distinctive optimization or ML substance — 3/10. The system combines specialized workflows, external models, retrieval, and validation. Historical evidence identifies conventional machine-learning work. No reproducible benchmark or published method establishes a distinctive optimization capability beyond this orchestration and application engineering. (2, 14)
Real-world constraint handling — 2/10. Search criteria can encode supplier characteristics, and source checks can enforce specific publication rules. These do not demonstrate joint treatment of procurement capacity, commercial terms, inventory availability, or delivery uncertainty. The product addresses information qualification rather than feasible supply chain plans. (9, 3)
Decision automation and labor productivity — 5/10. The sourcing trial documents reduced research time, and other workflows produce files or update connected applications. The remaining customer review and vendor tuning prevent treating those task gains as comprehensive autonomous decision-making. (7, 9)
Resilience under real operational complexity — 2/10. Routing, retries, and escalation address workflow execution. The public evidence does not quantify behavior when sources become stale, external models change, validation fails, or human review queues accumulate. Scaling the number of agents does not resolve these failure modes. (2, 3)
Product and architecture integrity
| Criterion | Score |
|---|---|
| Architectural coherence | 5/10 |
| System-boundary clarity | 5/10 |
| Security seriousness | 3/10 |
| Software parsimony and ownership cost | 3/10 |
| Compatibility with programmatic and agent-assisted operations | 5/10 |
Average: 4.2/10
Architectural coherence — 5/10. Specialized agents, coordination, validation, and task-specific interfaces form an intelligible application structure. The claim that models can change without affecting the rest of the pipeline ignores behavioral dependencies. The division into layers needs regression evidence, not just a diagram. (2)
System-boundary clarity — 5/10. The integrations page distinguishes source reads, CRM writes, calendar actions, and document exports. The sourcing workflow also clearly ends with a buyer’s choice. Credential scope and the technical boundary between generation, validation, and publication remain less explicit. (7, 9)
Security seriousness — 3/10. Deployment choices and retention commitments address relevant concerns. Public material does not establish tenant enforcement, privileged-access controls, prompt-injection containment, or task-specific authorization. These omissions are consequential for agents that read external content and act through enterprise accounts. (4, 19)
Software parsimony and ownership cost — 3/10. The historical architecture contains multiple persistent data systems, and current operations add orchestration and model-provider dependencies. Including maintenance in a base fee clarifies who performs some work; it does not show that the total stack is economical to maintain. (15, 5)
Compatibility with programmatic and agent-assisted operations — 5/10. Agent-to-agent handoffs, APIs, and business-system writes can remove manual transfers. Public documentation does not expose enough interface or authorization detail to independently inspect a complete workflow or safely replace one component. (2, 7)
Technical transparency
| Criterion | Score |
|---|---|
| Public technical documentation | 3/10 |
| Inspectability without vendor mediation | 2/10 |
| Portability and lock-in visibility | 4/10 |
| Implementation-method transparency | 4/10 |
| Evidence density behind technical claims | 3/10 |
Average: 3.2/10
Public technical documentation — 3/10. Current pages describe architecture, checks, integrations, and deployment. They provide no detailed API reference, validation algorithm, test protocol, or operational failure analysis. Historical hosting and engineering records supply more concrete stack information than the present product pages. (2, 3, 14, 15)
Inspectability without vendor mediation — 2/10. Buyers cannot reproduce the zero-error claims from a public dataset or runnable example. The supplier-search trial is inspectable as a reported experiment, but its query set, relevance assessments, and complete results are not supplied. A summarized outcome is not an independently repeatable evaluation. (3, 9)
Portability and lock-in visibility — 4/10. Editable InDesign output and customer-hosted deployment reduce some practical barriers. The terms reserve the platform, orchestration logic, and underlying agent methodologies to the vendor. Owning a report does not provide the ability to regenerate it without that machinery. (6, 7, 8)
Implementation-method transparency — 4/10. The company describes scoping, data integration, supervised testing, and subsequent operation, with continuing retuning included. The process does not explain acceptance-test coverage, rule versioning, rollback, or how customer-specific configuration remains reliable as its source systems change. (1, 5)
Evidence density behind technical claims — 3/10. The sourcing trial, cloud-provider account, and engineer’s work history establish identifiable activities. The strongest reliability and scale claims have substantially less supporting detail. A second validation agent and cited sources do not prove universal correctness. (3, 9, 14, 15)
Vendor seriousness
| Criterion | Score |
|---|---|
| Technical seriousness of public communication | 2/10 |
| Technology roadmap integrity | 4/10 |
| Conceptual sharpness | 3/10 |
| Economic accountability | 3/10 |
| Defensibility in an agentic-software world | 3/10 |
Average: 3.0/10
Technical seriousness of public communication — 2/10. Unqualified claims of zero error, no incorrect information, and unchanged behavior after model substitution exceed the disclosed methods. The public terms are more candid about the possibility of mistakes than the sales pages. (2, 3, 6, 11)
Technology roadmap integrity — 4/10. Research assistants, cloud infrastructure, and current workflows show sustained development. Earlier marketplace ambitions and broad AI promises lack a clear delivery account. The recent sourcing trial is more persuasive evidence of progress than continuity of the agent label. (15, 16, 17, 9)
Conceptual sharpness — 3/10. Separating generation from validation and maintaining source provenance are sensible design choices. Treating provenance as correctness and report production as validated intelligence weakens the reasoning built on those choices. The scalability account measures elapsed time without the associated resource cost. (3, 8, 12)
Economic accountability — 3/10. The base fee assigns integration and continuing tuning to the vendor. Mandatory customer verification and broad limitations on output responsibility narrow that commitment. Public evidence does not reconcile the promised transfer of operational risk with the customer’s remaining labor and exposure. (5, 6)
Defensibility in an agentic-software world — 3/10. Maintained enterprise integrations, task-specific validation, and customer configuration can be valuable. General model orchestration, company search, and report generation face increasingly accessible substitutes. The vendor has not demonstrated that its proprietary layer produces a sustained quality or cost advantage over those alternatives. (2, 5, 7)
Sources
Sources were reviewed on November 4, 2026. Historical architecture and engineering records are dated separately from current product claims.
[1] Current positioning and deployment process
AGENTS.inc: Homepage. Undated. Managed marketing, sales, and intelligence agents, customer claims, and implementation sequence.
AGENTS.inc: Platform overview. Undated. Agent fleet, coordination hub, model choices, and validation layer.
AGENTS.inc: Accuracy and Trust. Undated. SAP price checking, independent validation, provenance, escalation, and zero-error claims.
[4] Security and deployment boundaries
AGENTS.inc: Security and Compliance. Undated. Kubernetes, hosting choices, workspaces, and retention commitments.
[5] Pricing and maintenance responsibility
AGENTS.inc: Pricing. Undated. Base fee, ongoing retuning, maintenance, and task pricing.
[6] Output review and commercial obligations
AGENTS.inc: Terms of Service. Sections 5–6 cover output verification and consumption charges; later sections cover ownership and liability.
AGENTS.inc: Integrations. Undated. Source reads, CRM writes, calendar and email operations, and document output.
AGENTS.inc: Scaling and deployment. Undated. Managed and customer-hosted options, parallel-processing claims.
[9] Machu Picchu Foods sourcing trial
EU-LAC Digital Accelerator: Machu Picchu Foods and AGENTS.inc. September 24, 2026. Customer-defined trial targets, reported search results, and prospective expansion.
AGENTS.inc: Campaign management. Undated. Product-data ingestion, copy, layout, price checks, and claimed production-time improvement.
[11] Insurance deployment scope
AGENTS.inc: Insurance and Financial Services. Undated. Fink & Wagner AXA/Salesurance deployment, adjacent group-scale description, and perfect-accuracy claim.
AGENTS.inc: Market intelligence. Undated. Company mapping, external monitoring, and structured sector reports.
[13] Innovation scouting workflow
AGENTS.inc: Innovation scouting. Undated. Company discovery, fit assessment, monitoring, contacts, and presentation generation.
[14] Attributed engineering work
Dmitrii Kochetov: Public engineering CV. Undated document describing AGENTS.inc work in 2019 and July–November 2024. Python, scikit-learn, Kubernetes, ArgoCD, and Terraform are attributed to those roles.
[15] Historical cloud architecture
Scaleway: Agents.Inc. prepares to disrupt the AI knowledge market. Undated account concerning the 2021 launch. AWS migration, containerized agents, marketplace plans, and an architecture diagram naming the data services.
[16] Founder account of the early product
Sebastian Denef, OWN intelligence: This is our story. August 16, 2018. Collaboration software, the shift toward agents, initial framework, investment, and prototypes.
[17] Research assistants in 2019
Technology Park Berlin Adlershof: A pipeline full of ideas. March 1, 2019. Interview with the founders describing research agents, the shared board, and early trials.
AGENTS.inc: Imprint. AGENTS HQ GmbH, Berlin.
[19] Data processing and subprocessors
AGENTS.inc: Privacy Policy. Undated. Workspace data, public-source collection, model APIs, and hosting subprocessors.