Agents that act
inside the systems you run.
Blumcore is the runtime and the engineering practice under every Kit agent: a bounded loop, the right model for each step, one shared model of your business, and every action checked, approved and audited.
Bounded agent loop
Right model for each step
Approved, traced, audited
Starts with Blumcore Assist, a working prototype on your tenant in two weeks
An agent is a system,
not a prompt.
A demo agent is one prompt and one model. A production agent is a loop that plans, acts, observes and reflects, on the right model for each step, against one model of your business, inside budgets, permissions and approvals your teams set. Blumcore is that system, and the way we engineer on it.
Works in a loop, with limits
Plan, act, observe, reflect, with a cap on steps and clear stop conditions.
Routes every step
Small, fast models for simple steps; frontier models where reasoning matters.
Shares one model of the business
An ontology of your entities and relationships, so every agent reads the business the same way.
Stays inside your rules
Scoped tools, policy checks and human approval before anything consequential.
At a glance
Bounded agent loop
Any model provider
Runs air-gapped too
Runs every Kit agent
Right for you when
- You want agents that do work, not chat about it.
- Your data cannot leave your tenancy, or your network.
- You will not lock your agents to one model provider.
- You need to see what every run did, and what it cost.
One agent run,
step by step.
What happens between a task arriving and a record changing in your systems.
The task is read against the ontology. Retrieval pulls only the records and documents it needs, and memory adds what earlier runs learned.
You receive
Grounded context
Six layers,
one runtime.
Each layer is a working part of Blumcore and a habit of how we engineer agents on it.
Loop
The agent plans, acts, observes and reflects until the goal is met.
- Capped iterations
- Explicit stop conditions
- Hands back to a person when stuck
Routing
Each step goes to the model that fits it, and providers can change without a rebuild.
- Small, fast models for simple steps
- Frontier models for reasoning
- Azure OpenAI, OpenAI, Anthropic, Gemini or self-hosted
Context and cost
The model sees what the step needs and no more, so runs stay fast and spend stays predictable.
- Prompt caching
- Context compaction on long runs
- Retrieval instead of stuffing
- A budget on every run
Ontology and memory
Your business entities and their relationships, modelled once and shared by every agent.
- One semantic layer for all agents
- Short-term working memory
- Long-term memory across runs
Tools and policy
Agents act through MCP tools with scoped permissions, inside the rules you set.
- Least-privilege tool access
- Guardrails on inputs and outputs
- Human approval gates
Evals and observability
Every change is tested before release, and every run can be inspected after it.
- Evaluation harness and regression tests
- Traces, cost and latency per run
- An audit log of every action
Proof from Andhra Pradesh.
Government of Andhra Pradesh
Government · Tax administration
- Manual compliance paperwork cut through automated notice drafting
- 80%
- Dossier preparation time, through automated notice drafting
- Days → minutes
AI-Drafted GST Enforcement in Andhra Pradesh, Signed by an Officer
AI Tax Officer runs on the Blumcore runtime for the Commercial Taxes Department: the agents draft, and an officer reviews and signs every notice.
Agents that do the work,
not just talk about it.
Construction & AEC
RFIs, schedules and proposals drafted by agents that act inside Outlook and Primavera P6.
Know more about Construction & AECManufacturing
Agents that read SAP and plant records and raise the next action inside the systems already in use.
Know more about ManufacturingGovernment
Tax notices drafted on Blumcore and signed by officers, every step audited.
Know more about Government
Platform independent.
Same agents, any stack.
Blumcore keeps the agent apart from the model, the data platform and the surface, so any of them can change without a rebuild.
- Cloud, on-premise or air-gapped
- Weights and prompts stay inside your perimeter
- Every tool call permissioned and audited
- Systems of recordMCP-native connectors; ServiceNow and NetSuite too
SAP
Oracle
Salesforce
Dynamics 365
- Data platformGoverned data read where it already lives
Microsoft Fabric
Databricks
BigQuery
- Language modelsRouted per step; change provider without a rebuild
Azure OpenAI
OpenAI
Anthropic
Gemini- Self-hosted models
- Speech modelsFor agents that listen and speak
ElevenLabs
Azure AI Speech
Google Speech
- SurfacesWhere your teams meet the agent
Teams
Outlook- Web and API
Questions buyers ask.
What is Blumcore?
Blumcore is the Mandelbulb Kit runtime that runs AI agents inside your systems of record. It gives each agent a bounded loop to plan, act, observe and reflect, routes every step to the right model, grounds the work in one ontology of your business, and puts policy, human approval and an audit log around every action. Atlas AI, AEC Platform and the AI Tax Officer all run on it.
How does Blumcore choose which model to use?
Blumcore routes each step to the model that fits it. Simple steps such as classifying, extracting or formatting go to small, fast models; planning and hard reasoning go to a frontier model. Providers are swappable: Azure OpenAI, OpenAI, Anthropic, Gemini or a self-hosted model. Because the agent stays apart from the model, you can change provider, or add a new model, without rebuilding the agent.
How does Blumcore keep token cost and latency under control?
Blumcore controls cost by giving the model only what each step needs. It caches stable prompts, compacts context on long runs, and retrieves the relevant records instead of stuffing whole documents into the prompt. Every run carries a budget for steps and tokens, and the loop stops when it reaches one. Traces show cost and latency per run, so your teams can see where spend goes.
How does Blumcore keep agents safe and accountable?
Blumcore keeps agents inside rules you set. Agents act through MCP tools with scoped permissions, guardrails check inputs and outputs, and a person approves anything consequential before it lands in a system of record. Every change to an agent passes an evaluation harness and regression tests before release, and every action is logged, attributable and reversible. It runs in the cloud, on-premise or air-gapped.
How does a Blumcore engagement start?
A Blumcore engagement starts with Blumcore Assist, a working prototype on your tenant in two weeks. Bring the workflow you would hand to an agent first to a 30-minute working session, and we agree the systems it must read and act in, the entities it needs in the ontology, and where a person approves. The prototype runs against your own data and permissions, so the team judges it on real work.
Also in
the Kit.
Atlas AI
Inbox intelligence
Project mail triaged, commitments tracked and the next move proposed, with nothing sent unseen.
Know more about Atlas AIAEC Platform
Construction agents
RFIs, schedules, proposals and CVs drafted from the project record and signed off by your engineers.
Know more about AEC PlatformAI Tax Officer
Tax administration
Returns reconciled, risk scored and notices drafted for an officer to review and sign.
Know more about AI Tax Officer
Put one agent
to work on your systems.
Bring the workflow you would hand to an agent first to a 30-minute working session.