AlphaAgent OS

The hardest parts, pioneered.

A model and instructions are only the start. AlphaAgent OS brings ten proprietary components together to solve the hard operating problems around real agentic work.

Inherited operating guarantees

The complete operating system

Every agent · Every team · Every workflow
  • Apply your expertise

    Institutional knowledge becomes reusable intelligence.

    • Intelligence Graph
  • Share what works

    Move proven assets between teams and deployments.

    • Organisation Library
  • Carry real work

    Agents compute, coordinate and work through approved systems.

    • Kernel
    • Flows
    • Bridge
  • Remember and improve

    Context persists. Improvements stay deliberate.

    • Memory
    • Academy
  • Keep control active

    Decision rights and supervision remain in force.

    • Authority
    • Governor
  • Make outcomes defensible

    Sources, computation and decisions remain inspectable.

    • Lens

Your teams inherit the complete system and every advance we make to it, so you can focus on reinventing the business capabilities that run on it.

  • Intelligence Graph

Apply your expertise.

Intelligence Graphs turn your methods, policies, research and institutional knowledge into reusable specialist expertise. Give an out-of-the-box agent a specialist skill: the knowledge of what to do, how to do it and which evidence to use.

Make it a specialist.

A prompt directs.

Sets the role, objective and instructions for the task.

A knowledge base informs.

Makes relevant documents and information available to search.

An Intelligence Graph equips.

Brings methods, relationships and source evidence together so an agent can apply a specialist skill.

Build that expertise once. Reuse it across authorised agents.

Capable agent + specialist skill = specialist agent

Specialist skills in action

One agent. Specialist skills.

A motor claim needs damage assessment, policy judgement and clear communication. Follow one agent as it draws on a different Intelligence Graph for each job, carrying the claim and its findings through to a case ready for review.

ONE CLAIM. ONE AGENT.

INSURANCE CLAIM

A claim arrives.

A collision at a junction.

Illustrative example

The customer reports front-left damage. Photographs, a repair estimate and the motor policy accompany the claim.

  • Accident description
  • Vehicle photographs
  • Repair estimate
  • Customer policy

Impact assessor

Front-left collision. Retrieve the assessment method, hidden-damage checks and report requirements.

  • Map contact direction against the visible damage pattern. [1]
  • Check bumper mounts and headlamp supports for concealed damage. [2]
  • Separate observed, inferred and unconfirmed damage in the report. [2]
Damage assessment report
  • Front-left damage is consistent with the reported contact.
  • The bumper and headlamp require inspection against the repair estimate.
  • Concealed mounting damage remains unconfirmed. Request inspection evidence.

Policy & legal reviewer

Damage assessed; concealed damage unconfirmed. Retrieve the coverage-review method and decision checks.

  • Match the loss date and incident to the operative policy wording. [1]
  • Check exclusions, excess and outstanding obligations. [2]
  • Escalate uncertain interpretation; keep repair scope provisional. [2]
Coverage recommendation
  • The reported collision appears to fall within the accidental-damage provision.
  • Confirm the applicable excess and outstanding policy checks.
  • Hold the final repair scope pending inspection. Submit the recommendation for approval.

Customer communicator

Inspection pending; coverage not approved. Retrieve guidance for a clear, accurate customer update.

  • Explain what was reviewed and which findings remain unconfirmed. [1]
  • State the inspection needed and what happens next. [2]
  • Avoid a coverage or settlement promise before approval. [2]
Draft customer response
  • “We have reviewed your photographs and repair estimate.”
  • “An inspection is needed to confirm whether there is damage behind the bumper.”
  • “We will confirm the repair scope and policy position once those checks are complete.”

The intelligence engine

Fully managed

Intelligence Graphs are powered by AlphaAgent’s proprietary Augmented Multi-Resolution Property Graph (AMPG) mechanism, which organises knowledge and retrieves it in context. Three stages turn source material into expertise an agent can use.

  1. 01 / Embedding

    Build the graph.

    Add policies, procedures, research and model documentation. AlphaAgent turns them into structured, contextual Intelligence Graphs, with the whole document and its detail held together. Each build creates a version for agents to pin to.

  2. 02 / Indexing

    Connect the knowledge.

    Concepts, relationships and overlaps are mapped at multiple levels of detail. New documents and domains can be added without designing and maintaining a rigid knowledge model for each one.

  3. 03 / Retrieval

    Apply what matters.

    For each task, the system finds relevant knowledge and follows the relationships that reveal the evidence that matters. Agents receive specialist context with a traceable path back to its source.

During agent work

  1. Ground before planning

    Specialists begin with the organisation’s relevant knowledge before deciding how to approach the work.

  2. Stay within bounds

    Retrieval follows an explicit budget, keeping exploration controlled and repeatable.

  3. Trace every source

    The Thought Map records the path from the question to the source. Retrieval supplies evidence, not an authored conclusion.

Specialist analytical work

A volatility analyst draws on model documentation, research and house methods to assess a signal in its market context.

Specialist operational work

A claims investigator draws on policy wording, procedures and exceptions to assess a case against the relevant source clauses.

Build the expertise once. Give every authorised agent the skill to apply it.

See how Intelligence Graphs work

Reusable expertise

Reuse specialist expertise.

Prometheus-builtAvailable

Specialist Intelligence Graphs from Prometheus Research Labs.

AlphaAgent ships with a growing set of Intelligence Graphs for high-value financial services use cases across capital markets, insurance, banking and payments.

Customer-builtAvailable

Your own institutional knowledge.

Teams can build Intelligence Graphs from their own policies, procedures, research, model documentation, operating guidance and specialist methods.

Organisation LibraryAvailable

Share expertise across deployments.

Share a selected Intelligence Graph version to an organisation library. Authorised teams can pull a copy into their own Studio deployment.

Agents stay pinned to the Intelligence Graph version chosen for them until someone moves them. Share that version through the Organisation Library. Each team’s copy records its source and stays under its own deployment’s control.

See how libraries and sharing work

Use cases

Built for two kinds of specialist agents.

Intellectual agents

For specialist work that requires deep applied knowledge.

Quant researchers, volatility specialists, FX analysts, portfolio researchers and credit analysts need more than general language understanding. They need embedded knowledge, technical methods, mathematical context and the ability to apply them correctly.

  • Specialist domain knowledge

  • Mathematical and model context

  • Research workflows

  • Expert reasoning paths

Operational agents

For regulated work that requires policy, guidance and control.

KYC analysts, fraud investigators, compliance reviewers, claims handlers and underwriters need to navigate huge volumes of policy, regulation, procedure and exceptions.

  • Policy and regulation knowledge

  • Procedure and control mapping

  • Source-grounded decisions

  • Traceable escalation logic

  • Organisation Library

Share what works.

Turn one team’s proven work into a starting point for the next. Share a chosen version through the Organisation Library. Authorised teams pull a complete copy into their own Studio, then decide what runs.

Organisation Library

A workflow carries the whole capability.

Share the work with everything it depends on. The next team starts with a connected capability, ready to review and activate.

  • Agents
  • Workflows
  • Intelligence Graphs
  • Connectors
  • Environments

  1. Choose a version.

    Share a selected asset version from Studio to the Organisation Library. The version you choose is the one other teams receive.

  2. Pull the complete workflow.

    Authorised teams receive their own copy, with its agents, intelligence, connectors and environments already wired together.

  3. Activate deliberately.

    The workflow arrives as a draft. The receiving team reviews it and chooses when to activate it in its own Studio deployment.

  • Access belongs to the firm.

    An administrator creates each resource-type library and assigns people or Entra ID groups. Contributors share and pull. Readers pull.

  • Every copy has a source.

    Each copy records its origin and source version. Every share and pull stays on the record.

  • Updates remain a choice.

    When the source improves, teams choose Update my copy. They review and activate the workflow changes in their own deployment.

Expertise spreads. Each deployment stays in control.

Connector credentials move directly between the firm’s own deployments. They are never stored in the Organisation.

  • Kernel
  • Flows
  • Bridge

Carry real work.

Agents compute against real data, coordinate adaptive work and use approved systems. Substantial runs continue without step-by-step supervision.

Kernel

Runtime execution

Execute the work.

Coordinate specialists
Run individual agents, coordinated teams, tools and supervised specialist work through one durable runtime.
Compute the answer
Run Python and shell analysis against real datasets. Bring your own Python packages and Java libraries in custom execution environments, sandboxed on AWS Lambda inside your AWS account. Code reaches networks through attached connectors, within the configured boundary.
Continue long work
Checkpoint progress, reconnect people to the same run and notify them when work completes or needs attention.

The output comes from work that actually ran, not text that merely sounds calculated.

Flows

Governed orchestration

Keep judgement intact.

Combine adaptive steps
Move between agent work, document analysis, conditions, branches and human handoffs without reducing everything to a script.
Operate reliably
Apply schedules, retries, timeouts, approval points and structured outputs where the process requires them.
Pass the full work
Move complete datasets and files between authorised steps while keeping large payloads out of the model context.

The process becomes repeatable without flattening the judgement inside it.

Bridge

Approved connection

Reach approved systems.

Connect business systems
Connect SQL databases, REST APIs, MCP servers, AWS services and approved business tools.
Test before activation
Studio drafts a connection guide from the available schema or specification. Validate the connection, tools and expected schemas before activation.
Protect every credential
Keep credentials in the deployment’s secret store, never in a prompt. Resolve them only for authorised calls. Use short-lived AWS credentials where available.

Agents reach the systems they need without being given unrestricted access.

One request, coordinated expertise

One team. One outcome.

Complex work draws on several disciplines. A supervisor brings the right specialists together, coordinates their tasks and combines their findings into one evidenced result.

Your people set the objective and review the outcome. The team carries the investigation between them.

Illustrative research team

“What is driving this volatility?”

Supervisor coordinates
  • Macro
  • Volatility
  • Risk
One sourced assessment

Match control to the work

Choose what stays fixed.

Decide what agents can choose as they work and what the process must determine in advance. Both approaches operate within defined authority and retain evidence of what happened.

Dynamic teams

Follow the question.

A supervisor selects specialists as the work unfolds. New evidence can change the next task, the tools used and the expertise needed.

A research desk brings together macro, volatility and risk expertise as an investment question develops.

  1. Set the objective
  2. Adapt the team
  3. Review the evidence

Governed workflows

Define the operating process.

Set the sequence, decision points and approvals in advance. Specialists still apply judgement within each step, while the process stays controlled.

A claims operation coordinates damage assessment, policy review and customer communication, with approval where required.

  1. Define the steps
  2. Apply checks
  3. Approve the outcome

Chat becomes workflow

Make proven work repeatable.

Explore the work in Studio first. When the approach proves useful, ask Studio to turn the completed plan into a draft workflow. Review the steps and controls, reusing the specialist intelligence and connections already in place.

  1. 01

    Explore in Studio

    Work through a question in chat with specialist agents.

  2. 02

    Capture the process

    Ask Studio to turn the completed plan into a draft workflow.

  3. 03

    Set the controls

    Review the steps, approvals and exceptions. Save and activate a version.

  4. 04

    Run it again

    Run the active workflow on demand, on a schedule or from your systems through its API.

The process can run without someone prompting each step. People retain the decisions that matter.

See how work runs
  • Memory
  • Academy

Remember and improve.

Memory and Academy make continuity and controlled improvement part of the operating system. Useful context survives the exchange. Better-performing approaches are tested, versioned and promoted without silently changing live work.

Memory

Persistent context

Context survives the moment.

Retain relevant history
Carry forward the interactions, decisions and outcomes that remain useful to future work.
Keep work products usable
Preserve datasets, files and generated artefacts so later work can build on them instead of recreating them.
Support longer objectives
Let specialists return to an objective with the context needed to continue coherently.

Memory becomes part of the working environment, not a transcript that must be reread.

Academy

Controlled improvement

Improvement stays controlled.

Test alternatives
Compare instruction, agent and routing variants against representative work.
Measure the outcome
Evaluate quality, reliability, efficiency and the criteria that matter to the capability.
Promote deliberately
Version the result and let people decide when a better-performing alternative becomes active.

The system improves without silently rewriting how live work operates.

See how work persists
  • Authority
  • Governor

Keep control active.

Authority and Governor solve different control problems. Authority keeps access and decisions with people. Governor supervises agent behaviour while work runs. Together they let agents carry more responsibility without weakening accountability.

Authority

Human decision rights

People keep decision rights.

Define responsibility
Control who may access, direct, approve, change and inspect each part of the work.
Place human review
Require a decision where business judgement or accountability matters. API-triggered workflow approvals wait in Studio’s Inbox. The API can inspect them, but only a person can release them.
Understand dependencies
See what relies on a resource before changing or replacing it.

Agents carry the work without inheriting authority that belongs to people.

Governor

Runtime supervision

Keep work on course.

Watch the objective
At checkpoints, each agent justifies its next action. Governor reviews it against the step’s objective and grounded facts. Supervision is on by default.
Intervene proportionately
Stay quiet while work is sound. Correct drift, request clarification and return the agent to its Intelligence Graph when it needs source evidence.
Keep people involved
Let authorised users add, edit, withdraw or apply guidance before it becomes part of the run.

Supervision supports the work without replacing the person accountable for it.

Governed connections

Define how data works.

A connection carries instructions as well as access. Agents need to understand what a source means, how to use it and where their permission ends.

Bridge connects the systems. Authority defines access and decision rights. Governor supervises the work that uses them.

Portfolio review connection

Illustrative access and usage rules

Sources
Portfolio positions, mandate documents and approved market data.
Meaning
What each field represents and how positions relate to mandate limits.
Instructions
When to use the connection, what to request and which joins are valid.
Permissions
Approved roles, permitted data and limits on changes or external access.
Approvals
Actions that require an authorised person before they can proceed.

Control during execution

Keep authority in force.

Agents can investigate and prepare a recommendation without gaining permission to carry out every action they propose. Access, supervision and approval work together throughout the run.

  1. Set the authority.

    Define who can use the capability, what it may access and which decisions need approval.

  2. Supervise the work.

    Governor supervises every agent by default. It reviews each step against the objective and grounded facts, requesting correction or clarification when needed.

  3. Retain the decision.

    The authorised person reviews material decisions. Guidance, interventions and approvals remain with the work.

Illustrative portfolio review

Assess a mandate breach.

Read approved positions and limits. Investigate the exposure and prepare a recommendation.

Changing the portfolio requires approval.

The proposed action goes to an authorised person. Permission to investigate does not grant permission to execute a trade.

The work moves forward. Accountability stays clear.

Lens keeps the evidence, guidance and human decisions together so the outcome can be inspected afterwards.

Personal data is masked by default. Workflow settings can tighten redaction, never loosen it.

See Governor in action
  • Lens

Make outcomes defensible.

Lens builds the audit record as work happens. It keeps configuration, evidence, computation and material decisions together, so people can inspect and defend the outcome without reconstructing it afterwards.

Lens

Evidence and evaluation

Keep the record with the work.

Preserve what ran
Record the exact resource versions, configuration and workflow path behind the result.
Keep supporting evidence
Bring together citations, web sources, Thought Maps, executed code, datasets and step outputs.
Show material decisions
Retain Governor guidance, human interventions, completion times and usage beside the work. Attribute each run to a person or API key and each approval to the person who released it.

Teams can inspect, replay and defend how an outcome was reached.

From request to record

See what actually happened.

A defensible answer needs a visible path behind it: what the agent was asked, which data it used, what it executed and which controls shaped the result.

Lens retains that path as the run progresses, from the first request through the evidence and decisions kept with the outcome.

Execution trace
Illustrative run
  1. Request received
  2. Connector policy checked
  3. Context retrieved and joined
  4. Tools selected
  5. Generated code recorded
  6. Code and tool outputs captured
  7. Response and evidence stored
  8. Audit record saved

Traceability

Keep the full context.

Inspect the code that ran, the tools called and the data consumed. See how records were joined and which evidence supported the response, alongside the guidance and approvals that affected it.

Runtime telemetry ledger

Illustrative record · retained inside your boundary

Attribution
The person or API key responsible for the run, and the person who released each approval.
Generated code
Generated and executed code, inputs, runtime version and execution result.
Tool calls
Requests, responses, latency, errors and the policy checks applied.
Data use
Connectors, documents, fields, rows and joins used during the run.
Agent responses
Intermediate outputs, the final answer and supporting evidence.
Review and decisions
Evaluations, Governor guidance, human interventions and approvals.

Evidence in use

Inspect. Evaluate. Defend.

Use the record to understand performance, assess quality and account for material decisions.

Observe

Inspect the run.

Follow agent activity, tool calls, data use and computation. See the outputs, timings and errors behind the result.

Traces · Logs · Metrics · Code outputs

Evaluate

Test the standard.

Assess agent behaviour and output quality against defined criteria. Compare results and identify regressions before and after deployment.

Test suites · Quality scoring · Human review

Audit

Defend the outcome.

Review the sources, resource versions, workflow path and approvals together. Trace the result back to the work that produced it.

Provenance · Replay · Data lineage · Approvals

Deployment boundary

Your AWS. Private access.

AlphaAgent runs in AWS accounts you own and connects directly to your AWS Transit Gateway, with zero access from the public internet. Your agents, workflows, intelligence and results stay inside your deployment boundary.

Customer-controlled boundary

Your AWS accounts

Private access through AWS Transit Gateway

Only network ranges you approve can reach AlphaAgent, with sign-in through your Entra ID and MFA. Your network, DNS, identity and firewall policies control access.

AlphaAgent Studio

AlphaAgent OS

AlphaAgent Organisations

Private connections
  • Amazon Bedrock
  • Your data
  • Your systems

Amazon Bedrock and AWS services use private VPC endpoints. Your prompts, research and results never cross the public internet. Agents can reach only approved network ranges and ports, with read-only SQL access.

Customer work stays here
  • Agents and workflows
  • Specialist intelligence
  • Conversations and documents
  • Data and credentials
  • Results and evidence
Outbound onlySigned licence heartbeats and usage counts

Outbound traffic follows your firewall, allow-list, inspection and logging policies.

Private deployment guide

Architecture diagrams and network checklist

Start with one capability

Map the whole system.

Bring us a piece of work that matters. We will map the specialist intelligence, execution, connections, continuity, authority and evidence required to run it properly.

Talk to usRead the OS docs