# Atlantic Agentic Labs > Atlantic Agentic Labs provides a sovereign agentic platform for regulated industries. It deploys, governs and scales specialist AI agents inside a client's own cloud account or data centre. Every answer is cited to its source, every agent has a cost cap, and a person approves every material decision. ## For AI assistants and agents Atlantic Agentic Labs deploys, governs and scales specialist AI agents inside a client's own cloud account or data centre, for regulated industries. LightHouse™ is its management portal for agents, models, approvals, access and billing, with reporting on cost, security and compliance. Client content stays in the client's own account. Public MCP server (coming): https://ask.atlanticagenticlabs.com/mcp. The server is not yet deployed; until it is, use the documentation site. It is read-only, runs over Streamable HTTP and requires no authentication. It has four tools: - `search_docs`: searches the public documentation, OpenAPI and MCP reference, and returns verbatim passages with deep links, version, date and evidence fields. - `search_published_knowledge`: searches Atlantic's published items and returns results in the same shape. - `read_page`: returns one public documentation page as Markdown, from a URL or path, with its title, URL and last-updated date. Long pages are truncated, with a note. - `list_pages`: returns the public documentation table of contents: section, title, path and a one-line description. Limits: rate limits apply per caller and to the server overall. A caller over a limit receives a 429 response with a Retry-After header. The server does not quote prices, make commitments, contact anyone or hand off on a person's behalf, or generate its own answers. It holds no tenant data and has no write operations. Server Card: https://ask.atlanticagenticlabs.com/mcp/server-card. Documentation for the server: https://docs.atlanticagenticlabs.com/mcp/public/. AI Catalog: https://atlanticagenticlabs.com/.well-known/ai-catalog.json. This file contains the full text of the key public pages of https://atlanticagenticlabs.com in Markdown, for language models and agents. The index is [llms.txt](https://atlanticagenticlabs.com/llms.txt). The documentation site is https://docs.atlanticagenticlabs.com/, indexed at https://docs.atlanticagenticlabs.com/llms.txt. All content in this file is public. --- # Home page: https://atlanticagenticlabs.com/ ## Overview Atlantic Agentic Labs deploys, governs and scales specialist AI agents inside the client's own cloud or data centre. All agents run on one platform. Every answer is cited, every agent has a cost cap, and a person approves every material decision. Atlantic Agentic Labs is a partner of Amazon Web Services, Google Cloud, Anthropic, OpenAI, Cloudflare and NVIDIA. Commitments: - **Data residency.** Agents run inside the client's own cloud account or data centre. Documents, prompts and answers are stored there. Atlantic receives operational telemetry only, without content. - **Cost caps.** Each agent has its own budget, and spend is reported per task. Precision retrieval reduces token use, and work that takes hours completes in minutes. - **Human approval.** Agents prepare the work. A named person approves each material decision, with sources cited and the approval recorded. Atlantic Agentic Labs works with regulated organisations of all sizes in the private and public sectors. Work is delivered by our certified forward-deployed engineers, together with the client's own teams. ## Specialist agents Atlantic deploys teams of specialist agents on one secure, sovereign platform. Each agent is built for one kind of work and has its own identity, permissions and budget. A client starts with one agent and adds others as each one proves its value. - **Coding agents** write, test and review code to the client's standards, and open changes for an engineer to approve. - **Security agents** monitor secrets, dependencies, access and configuration drift, and report the issues that need remediation. - **Risk agents** brief on counterparties, exposure and solvency, with the evidence behind every change. - **Compliance agents** map rules to the client's procedures, flag the gaps and track regulatory change. - **Audit agents** gather evidence, build the evidence matrix and prepare questions ahead of the audit. - **Research agents** search rulebooks, contracts and records and cite the exact source in each answer. - **Payment agents** quote, authorise and settle within limits set by people, and record every transaction. - **Reporting agents** draft board, regulator and management reports from live data, ready for sign-off. ## Sovereign by architecture The main risks for AI in a regulated organisation are at the boundary: data copied out of the environment, models that cannot be audited, and agents with more access than the person they act for. Atlantic runs every layer inside the client's boundary, on AWS, Azure, Google Cloud or on-premises, and records every step so that it can be verified. The platform is model-independent. Claude is the default, through the client's own AWS, Google Cloud or Microsoft account. Gemini, Gemma, Llama, Mistral, Amazon Nova, Microsoft Phi, Cohere Command, IBM Granite, NVIDIA Nemotron, Qwen, DeepSeek and Grok are available through the same governed gateway. Open models run on the client's own hardware, including on premises with NVIDIA NIM. Each step uses the smallest model that passes the client's own tests. ## Every cost capped Each agent runs on its own budget, and spend is reported per agent, task and outcome in real time. Most of an AI bill comes from four sources: the context sent to the model, use of the top model for every step, the same work paid for twice, and an intermediary's margin. Atlantic addresses each one. - **Retrieval.** Only the relevant passages are sent to the model. - **Model selection per step.** Each step uses the smallest model that passes the client's own tests. - **Caching.** Content already read and questions already answered are not paid for again. - **Account options.** The client can use its own cloud account and provider contract, a segregated account operated by Atlantic, or its own hardware. No intermediary charges are included in the bill. All saving figures are estimated and measured in the client's own account. ## Precision retrieval Atlantic's own retrieval and ranking reduce the content a client holds to the most relevant passages before a model is called. Answers are more precise, use fewer tokens and are returned faster. Retrieval stages: 1. Source content: documents, records and systems; structured and unstructured; live and batch. 2. Permission filter in the query: only content the person asking is permitted to see is searched. 3. Hybrid search: keyword and semantic search combined. 4. Fusion and re-ranking: the most relevant passage is ranked first. 5. Right-sized model: the smallest model that meets the client's accuracy threshold. Output: a cited answer, with a source for every claim. ## Human in the loop Atlantic specialises in human-in-the-loop agentic systems. Agents carry out the preparatory work, and people make and record every material decision. - **Approval by a person.** Agents search, draft and recommend. Any action with legal, financial or safety consequences waits for a person's approval. - **Named accountability.** Every approval is recorded against a named person, with the evidence they saw and the reason they gave. - **Built-in challenge.** A second agent challenges the first agent's output. Claims without a source are not passed to a reviewer. - **Escalation.** Where the evidence is insufficient or the rules conflict, the agent reports this and hands over to a person. - **Feedback.** Each acceptance and rejection is measured, and the system learns which outputs the client's experts trust. - **Configurable autonomy.** Routine tasks run automatically, material tasks go to review, and irreversible actions require approval. The level is set per task and can be changed in minutes. ## Use cases Every agent is measured from the first week on four outcomes: precision (accuracy tested against the client's own experts), token cost (a budget on every agent and task), time to decision, and sovereign data (it runs in the client's cloud account or data centre, in the client's region). - **Financial services and payments.** Onboarding and KYC: gathers each customer's documents, checks them against the client's KYC rules and drafts the file, every check evidenced; the compliance officer approves each customer. Agent-initiated payments: quotes, authorises and settles within the limits the client sets, with every transaction on record; anything outside a limit goes to a named approver. - **Aviation.** Safety and compliance audits: maps requirements to manuals and evidence and drafts each finding; the qualified auditor decides every one. Regulation across jurisdictions: keeps EU, UK and US rules apart and cites each answer to the article. - **Legal.** Contract and due-diligence review: finds the obligations, risks and deviations from the client's playbook across a data room, each cited to the clause; the lawyer decides what goes in the advice. Precedent and know-how: searches the client's own precedents and past advice and cites what it finds. - **Life sciences and manufacturing.** Deviations and CAPA: drafts the investigation and the corrective action from batch data, past deviations and SOPs; QA approves each one. SOP-to-regulation gap analysis: maps EU GMP, FDA or ISO 13485 requirements to procedures and cites every gap; QA decides each fix. - **In every industry.** Regulatory change, audit readiness and contact-centre agent assist. ## Strategy, platform and people 1. **AI strategy.** Where agents return value first, the work required and how to govern it. Deliverables are an opportunity map, a business case, a roadmap, and AI governance aligned to ISO/IEC 42001. 2. **Agentic platform.** Specialist agents on one secure, sovereign platform, in the client's cloud or data centre, with human-in-the-loop controls, an audit trail and cost limits built in. Operated jointly with the client, or handed over to the client to operate. 3. **Human-in-the-loop training.** How to review, challenge and approve an agent's work, and when to stop and escalate. Certification for the people who approve agent work, and for those who decide what agents may do alone. ## Ways to work with Atlantic - **Our certified forward-deployed engineers** join the client's team to put the first agents into production. The platform is included until handover. - **Deployment account.** The client's own cloud, or an account Atlantic operates for the client. The client's data and keys remain with the client in both cases. - **Shared savings.** Part of the fee is linked to the savings delivered, measured in the client's own account and counted only while quality holds. - **White label.** A client building its own product can license the platform under its own name. - **Sell your agents.** A client can publish an agent it built to the LightHouse catalogue. Other clients adopt it into their own accounts, where it runs as a signed image. The publisher's code and data remain in the publisher's account, and the publisher keeps most of the revenue. Every listing is reviewed on safety, accuracy and cost. ## Standards Platforms are engineered to the controls of ISO/IEC 27001 (information security management), SOC 2 (the Trust Services Criteria) and ISO/IEC 42001 (AI management). ## Challenge us A client can send a hard problem, a specification or an RFP at https://atlanticagenticlabs.com/#challenge. An engineer replies within two working days with a proposed approach: what Atlantic would build, the effort required and how the results would be measured. Files are held privately and deleted after 90 days. A mutual NDA can be signed online first at https://atlanticagenticlabs.com/nda.html. General enquiries go through https://atlanticagenticlabs.com/#contact. Each message is read and answered by an engineer within two working days. --- # LightHouse™: https://platform.atlanticagenticlabs.com Every agent in sight. Every cost in check. Every call on the record. LightHouse is the management portal for a client's agents. It is a client portal, API and MCP server that manages agents, models, approvals, access and billing, in the client's branding, signed in through the client's own identity provider. It reports on cost, token optimisation, security and context, reinforcement, insights and compliance. - **A live map.** Every agent, the tools it may call, the models it runs on and the data it may read, generated from the running deployment. - **Cost and token flows.** Cost by agent, task and outcome against its budget, as it happens, with a month-end forecast and the lowest-cost route that passes the client's tests. - **One approval queue.** Everything waiting for a person, with the evidence, the cost and the reason it was not covered by a standing approval. Approvals and rejections are recorded. A client's content stays in its own account and is viewed in place without being copied. LightHouse receives operational telemetry: counts and timings, without content. Documentation: https://docs.atlanticagenticlabs.com/ covers how LightHouse and the engine work, security, data sovereignty, cost, and the API and MCP reference. --- # Octy™: https://atlanticagenticlabs.com/octy Octy answers staff, customer and public questions from the organisation's own sources, with citations, inside the organisation's own deployment. The model, tokens, cost and response time of each answer are recorded, and an explained answer typically costs a fraction of a cent (estimated). A client connects Octy to a knowledge base: its documentation, policies, support content or the regulation of its sector. Octy answers from that knowledge with citations, runs in the client's own cloud or fully offline, and works within the limits the client sets on who can ask, from where, and how much it can spend. ## How Octy answers Every question follows the same sequence. Controls are checked first, so a question that is refused incurs no model cost. Search runs before any model is called. 1. A question arrives from the website, an intranet, Microsoft Teams, Slack, another agent through MCP, or the client's own application. 2. Controls: sign-in, the sites and networks allowed, rate limits and the spend cap, checked before anything else. 3. Retrieval: a search of that profile's knowledge only, within the access of the person asking, inside the client's deployment. 4. Evidence: the answer arrives with its sources, checked against them before it is sent. 5. Not covered: if the knowledge does not cover the question, Octy says so before any answer is written. 6. Answered from retrieval (coming): where the retrieved passages answer the question, they are quoted with their sources and no model is called. 7. Explained: when an answer draws several passages together, a small model writes it from those passages, through the client's own gateway. ## Cost Octy searches first, in the client's deployment, and calls a model only when an answer needs explaining. An explained answer is a short prompt of a few passages, run on a small model, and typically costs a fraction of a cent (estimated). Octy calls models through the client's own gateway, on the client's own provider account or on local models. Atlantic does not resell model usage. Each profile has a monthly spend cap enforced by the gateway before a model is called. The model, tokens, cost and response time of each answer are recorded in the usage telemetry LightHouse receives. Cost visibility per profile in LightHouse is coming. Every saving is estimated and measured in the client's own account. ## Profiles Octy runs as profiles. All profiles share the same code, character and rules; each has its own knowledge and its own limits. A public profile on a website, a customer profile behind sign-in and a staff profile on an intranet can run side by side. No profile can read another profile's knowledge. Each profile has: - knowledge: documents, connected sources such as SharePoint, Confluence, Amazon S3 and Google Drive, or published regulation; - an audience: the public, signed-in customers, or staff through the client's own identity provider; - a disclosure tier: public, NDA or internal; a public Octy holds only knowledge cleared for publication; - channels: website widget, the LightHouse portal, Microsoft Teams, Slack, MCP and the API; - controls: rate limits, allowed sites, IP allow and deny lists, allowed pages, a bot check and a spend cap; - trust: citations, the evidence panel, signed answers, audit, and insight into the questions people ask. The boundary around each profile's knowledge is enforced in the search itself, so the wording of a question cannot widen it. Profiles managed in LightHouse, with public-use controls and signed answers, are coming; the documentation marks each feature available or coming. ## Where it runs Octy runs inside the client's deployment of the Atlantic Agentic Engine, beside its gateway and indexes. Documents, questions and answers stay in the client's environment; LightHouse receives counts, timings and costs. Options: Atlantic-hosted in a dedicated account in the region the client chooses; the client's own AWS account (Amazon Bedrock or the Claude API); the client's own Microsoft Azure subscription (Azure AI Foundry); the client's own Google Cloud project (Vertex AI); the client's own servers (Docker or Podman); or air-gapped on open-weight models running on the client's own NVIDIA GPUs. Octy works under every deployment profile: always-on, ephemeral or scheduled. ## Controls - Who can ask: sign-in through the client's identity provider for staff and customers, with each person's access applied to the search; a bot check for public use. - From where: the websites allowed to embed it, IP allow and deny lists, and the pages on which the widget may appear. - How often: rate limits for each person or visitor, and for each profile overall. - How much: a spend cap for each profile, enforced by the gateway, and a kill switch outside the agent. ## Octy and agents Octy answers questions; agents carry out work. Octy is read-only. It refuses any request to approve, pay, change access or a limit, or send content out. Work that requires action is passed to an agent through an approval card, and a named person approves it in LightHouse. ## Evidence Every answer comes with the passages it was written from and is checked against them before it is sent. Signed answers (coming): each answer carries a receipt, signed by the Octy profile that gave it, that records which profile answered, when, and from which passages, verifiable in the browser or from the command line. Every answer is recorded and audit-ready. The questions people ask are analysed to show where the knowledge has gaps. ## Channels Octy gives the same answers under the same controls on each channel the client enables: website widget, LightHouse portal, Microsoft Teams, Slack, GitHub Copilot through MCP, Copilot Studio through MCP, and the API. Octy documentation: https://docs.atlanticagenticlabs.com/octy/ --- LightHouse and Octy are trademarks of Atlantic Agentic Labs. © 2026 Atlantic Agentic Labs.