Arcesium Intelligence: Control Without Compromise

Summary

The latest Arcesium Intelligence release adds Deterministic Execution with multi-write support, agent identity, the Opterra® MCP, scale to 300 concurrent jobs, model selection, role-based access control, and SDLC for applications, giving investment operations teams the governance, predictability, and control to deploy agentic AI across real production workflows.

When Agents Do More, Control Matters More

Every investment operations team that has begun deploying AI agents eventually confronts the same tension. The more capable an agent becomes, the harder it is to predict what it will do. Capability and control are usually treated as a trade-off: give an agent more to do, and you give up certainty about how it does it.

For investment firms, that trade-off is intolerable. In investment operations, error tolerance is near zero. A misrouted trade, an unauthorized write, or an agent acting outside its intended scope is not a minor inconvenience. It is a control failure with real consequences.

The latest Arcesium Intelligence release is built on the opposite view. Predictability is what makes capability usable. Each new feature in Agent Studio removes a reason to hold agents back from more operational work. Not by loosening control, but by deepening it. Teams can now decide precisely who builds agents, what agents are allowed to do, how they execute, and what they write to the platform.

Our last Arcesium Intelligence release introduced a production-grade agent harness with version control, human-in-the-loop approvals, scheduling, reusable skills, and operational storage. introduced a production-grade agent harness for Agent Studio, with version control, human-in-the-loop approvals, scheduling, reusable skills, and operational storage.

This release takes the next step: the controls let that agents take on real operational work at scale.

1. Deterministic Execution: No LLM at Runtime, When You Need It

Some workflows require judgment, and LLM-based orchestration is the right tool for those. But other workflows follow known behavior, where the result should be the same every time. Running those through an LLM introduces unnecessary variability, cost, and risk, especially for writes to the platform.

What’s new: Agents can now run as fully deterministic Python scripts, with no LLM involved at runtime and no token costs. Deterministic Execution works alongside the existing LLM-based orchestration, including complex workflows made of agents of agents.

That gives teams a choice about where each part of a workflow sits:

  • Where a workflow calls for judgment, LLM orchestration remains the right tool.
  • Where behavior is known and the result should be the same every time, Deterministic Execution makes it fully predictable.

Deterministic Execution is also an option for firms with strict policies against AI at runtime. Cost is a secondary benefit; the primary one is determinism.

Multi-write support: Deterministic Execution is required for writes. Agent Studio now supports multiple writes in a single workflow, executed sequentially, with optional human-in-the-loop approvals. Because writes run without an LLM, teams know exactly what will be written to the platform.

Why it matters: Teams can run multi-step writes to Opterra with full predictability and optional human oversight at each step. That is the capability that makes agents useful for real operational work, with the control that makes it safe.

2. Agent Identity: Control What Agents Can Do

An agent’s access should be a deliberate decision. For operations teams, the key question is what each agent is allowed to do.

What’s new: With agent identity, users can now create distinct identities for their agents, each defining what an agent is allowed to do based on a subset of the user’s own permissions. For example, a user with read and write access can create a read-only identity and run an agent under it.

Agents act as their identity: restricted permissions, honored on every run.

Why it matters: Agent identity gives organizations a clear answer to a key control question: what is this agent allowed to do? Access becomes a deliberate, granular, decision.

3. Opterra MCP: Arcesium Data, Available to Your Enterprise AI Tools

Investment firms are increasingly building and deploying their own AI tools. They need those tools to work with the operational platforms where investment data lives.

The Opterra MCP launches with 20 tools, enabling investment firms to connect any enterprise AI tool to Opterra and use natural language to orchestrate their investment operations workflows. The open MCP layer means your AI initiatives connect to Arcesium data without closed-system lock-in.

Why it matters: Your teams can access Arcesium data from the enterprise AI tools they already use and trust, using natural language.

4. Scale and Resiliency: 300 Concurrent Jobs, With Guardrails

Operational workflows do not run one at a time. As agents take on more work, the platform needs to handle real production volume without breaking, and without runaway costs.

What’s new: Agent Studio now supports 300 jobs running concurrently. It also sets limits on automated activity to prevent excessive calls. Default limits apply to API calls per session and tool calls per agent, and thresholds can be configured based on client requirements.

Why it matters: Scale without guardrails is risk. Scale with guardrails is capacity. Your teams can run production-grade agent workflows at volume, knowing the platform will hold, and knowing there are limits in place to prevent excessive automated activity.

5. Model Selection: Choose the Model for Each Agent

Different workflows have different requirements. A workflow that calls for nuanced judgment may need a more capable model; a workflow that follows a known pattern may not.

What’s new: Users can now choose which model an agent uses directly in Agent Studio, and change it as requirements evolve.

Why it matters: The right model for the right workflow, controlled by your team, adjustable as needs change. That means better outcomes where judgment matters and lower costs where it does not.

6. Role-Based Access Control: Decide Who Can Do What

Organizations managing sensitive investment workflows need to delegate responsibly, and that takes granular control over who can do what.

What’s new: Role-based access control introduces roles and groups. Four roles define what a person can do:

  • Admin: Super access across Agent Studio.
  • Artifact author: Creates and edits agents.
  • Viewer and executor: Views and runs agents, but cannot modify them.
  • Publisher: Approves and activates agents.

Agents are now also organized into groups, to enable even more granular access management. Administrators create a group, assign members, and set permissions for that group: read, execute, or edit. A member’s role determines what they can do within it.

Why it matters: Teams can control which users can access which agents and perform which actions.

7. SDLC for Applications: Manage Apps the Way You Manage Agents

Applications in Agent Studio give investment operations teams a place to run end-to-end workflows and invoke agents from within them.

What’s new: Applications now follow the same SDLC as agents, including version control and auditability, so what teams build as applications is managed the way their agents are.

One example: a workflow launched from an application takes a PDF file, extracts the data, passes it through agents for parsing and mapping, and writes a journal entry back to Opterra.

Why it matters: Your teams can initiate complete operational workflows from applications, with the same control and lifecycle management they apply to agents.

Predictable by Design, Capable in Execution

Each of these capabilities answers a question teams ask before they give agents more operational work. Together, they mean:

  • Deterministic Execution: Builders can run agents as deterministic Python scripts with no LLM at runtime, including multi-step writes with optional human-in-the-loop approvals, so teams know exactly what will be written.
  • Agent identity: Agents run under identities with restricted permissions, honored on every run.
  • Opterra MCP: Any enterprise AI tool can access Arcesium data through the Opterra MCP, with 20 tools at launch.
  • Concurrent jobs: Agent Studio supports 300 concurrent jobs, with configurable limits on automated activity.
  • Model selection: Users choose the model for each agent.
  • Access control: Granular control over who builds, runs, and publishes agents through roles and groups.
  • SDLC for applications: Applications follow the same SDLC as agents.

The agent is only one part of the workflow. The controls around it determine who can build it, what it can do, and how it writes. Arcesium Intelligence brings those controls together in Agent Studio, helping teams give agents more to do while keeping control of what they are allowed to do.

Getting Started

These capabilities are available now to teams building on Arcesium Intelligence.

For guidance on configuring roles, groups, and agent identities, or on setting up Deterministic Execution and human-in-the-loop approvals for a specific workflow, contact your Arcesium relationship team.

Teams that have not yet started building with Arcesium Intelligence can explore the current library of agents, skills, and applications to identify use cases that could apply to their own investment operations workflows.

Request a Demo | Learn more about Arcesium Intelligence

Authored By

Vera Shulgina

Vera is responsible for Arcesium's data strategy with a focus on driving value for clients through data solutions and data partner integrations.

Share This Post

subscribe today

No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.