Sentienta Blog

Category: Governance

  • From AI Capability to Agent Management

    For the past year or two, enterprise AI has largely been an experiment in capability.

    Can AI summarize a contract, analyze a spreadsheet, research a market, answer a customer, or write production code? Companies tested models against work people already performed and tried to determine whether the results were good enough to matter.

    That period is ending much faster than many organizations expected. The question is no longer simply whether AI can do useful work. It is how companies will manage dozens, or hundreds of agents performing work simultaneously.

    Cost is making this transition particularly visible. An agent does not necessarily generate one answer and stop. It may inspect context, develop a plan, search for evidence, call tools, evaluate the result, revise its approach, and continue until the work is complete or human judgment is required. Several agents collaborating on one objective can multiply that activity.

    This helps explain two important industry trends. Lower-cost and open-weight models, including models from DeepSeek, Qwen, and Kimi, are giving businesses more alternatives to expensive frontier models. At the same time, NVIDIA has put its Groq 3 LPX inference accelerator into full production, targeting the speed and responsiveness required by agentic workloads.

    NVIDIA describes the system in terms of “interactivity,” meaning how quickly an agent can generate tokens and complete each step of its work. Groq 3 LPX is designed for workloads involving long contexts, verification, tool calls, and multistep reasoning. In one benchmark cited by NVIDIA, it generated approximately 3,400 output tokens per second while running a 31-billion-parameter model with a 100,000-token context. NVIDIA’s announcement

    The technology industry is preparing for agents to perform much more work. Enterprises now have to prepare to manage that work.

    The Problem: Cost Reveals a Larger Management Challenge

    Model price is an inadequate way to measure the cost of agent work.

    A less expensive model may require more attempts, more detailed instructions, and more human correction. A frontier model may cost more per call but complete a difficult assignment correctly on its first attempt. An assignment may also consume web research, tool calls, external services, parallel agents, retries, human review, and coordination with other work.

    The meaningful unit is therefore not the token or the individual model call. It is the completed assignment.

    Once companies look at the whole assignment, a much broader management problem comes into view. An organization has to determine which work should be assigned to an agent, which model is appropriate, how much time and effort the assignment justifies, and whether the result compares favorably with work performed by a knowledgeable person.

    It must also decide which resources each agent needs. A legal assignment may require access to contracts in Google Drive, while an engineering assignment may require GitHub and Jira. Giving every agent access to every system is unsafe, but denying necessary information makes the agent ineffective.

    As more agents operate simultaneously, their work must also be coordinated. Two agents may unknowingly solve the same problem, make conflicting changes, or depend on one another without recognizing the dependency. The output of one agent may need to become evidence for another, and a manager may need to resolve disagreements before downstream work proceeds.

    Some decisions can be delegated. Others affect customers, production systems, contracts, finances, or public communications and should remain subject to human approval. The manager needs to see those consequential decisions without being asked to supervise every routine action.

    Faster inference makes these problems more urgent. Agents work through repeated cycles of observation, decision, action, and evaluation. Reducing the latency of each cycle makes longer and more complex assignments practical. It also allows organizations to run more agents and initiate more work. NVIDIA’s investment in specialized inference infrastructure is therefore not just a story about faster hardware. It is evidence that agent work is becoming a production workload. NVIDIA’s technical overview

    The Solution: Manage the Assignment

    Companies need to manage agents at the level of the assignment rather than the individual model call.

    For each assignment, the organization should be able to understand who or what is responsible, which model and resources are being used, how the work relates to other assignments, what decisions require attention, and whether the result justified the total cost and human effort.

    This does not mean exposing every prompt, inference, or tool call to a manager. That would replace productive work with constant supervision. Routine execution should remain in the background. The management layer should surface progress, important findings, dependencies, blockers, approval requests, and outcomes.

    Model selection becomes one part of this process. Lightweight models may handle monitoring, classification, extraction, and routing. More capable models may be appropriate for difficult reasoning or consequential decisions. Some assignments may begin with an inexpensive model and escalate only when uncertainty is high. Others may benefit from several agents independently examining a problem before their conclusions are compared.

    The objective is not to identify one model that should perform every task. It is to choose an appropriate combination of agents, models, information, tools, and human judgment for the work at hand.

    This is the management layer Sentienta is working to provide. In a Sentienta Workroom, people and agents can collaborate around shared objectives while retaining different responsibilities, knowledge, models, and access to business systems. Workflows support the scheduled or event-driven activity needed to fulfill them.

    The Workroom preserves the management context above that activity. It allows consequential progress, dependencies, decisions, and outcomes to remain visible without reproducing every operational detail. Permissions and approval levels constrain what agents may access or change, while activity records support accountability.

    Sentienta does not yet provide complete assignment-level cost accounting, automatic model routing, or a comprehensive comparison of agent and human performance. Those capabilities will become important as agent deployment grows. The immediate foundation is the ability to organize responsibilities, coordinate work, govern actions, and keep human judgment focused on the decisions that matter.

    Conclusion

    Enterprise AI is moving from testing what individual models can do to managing agents as a workforce.

    The rapid growth of lower-cost models and specialized inference infrastructure will make many more agent assignments economically practical. That will increase the importance of deciding what work agents should perform, how their efforts should be coordinated, which resources and authority they should receive, and whether their results justify their cost.

    The companies that benefit most will not simply be those that deploy the most agents or purchase the cheapest inference. They will be the ones that learn to manage agent work effectively.

    The capability question is being answered. The management question is just beginning.

  • Sentienta Connected Services

    Giving Every Agent the Access Its Role Requires

    A cross-functional team may work together on the same problem, but that does not mean every person, or every agent, should have access to the same information.

    As agents become part of everyday work, organizations are unlikely to rely on one universal assistant that knows everything and can access everything. Legal, engineering, sales, finance, and HR each work from different systems, different obligations, and different kinds of evidence.

    Connected Services are Sentienta’s answer to that problem. They let users authorize agents to work with the systems where team knowledge already lives, including Google Drive, Gmail, Google Calendar, Slack, GitHub, Jira, and local resources through Sentienta Bridge, without turning that access into a universal permission for every agent in the Workroom.

    In Enterprise AI Needs a Governance Layer, we argued that agent governance begins with access: which agents can reach which organizational systems? In How Sentienta Governs Agent Actions, we described the governed path from a request to an external action.

    Connected Services put those ideas into practice. They allow teams to bring specialist agents and their knowledge sources into a shared Workroom while preserving boundaries around information, authority, and action.

    Access Should Follow Responsibility

    Connected Services let users authorize Sentienta to connect with the systems where their work already lives. The cloud services use supported authorization flows and service APIs. Sentienta Bridge provides a governed path to locally available capabilities such as OpenClaw and selected files.

    Connecting a service does not mean every agent gains access to it. Sentienta treats service access as an agent capability, not as a universal capability of the platform. A user connects a service and then determines which agents may use the relevant capabilities.

    For example, a legal agent might be enabled to access contracts in Google Drive, while an engineering agent is enabled to access GitHub and Jira. An HR agent may need personnel documents and scheduling information. These agents can participate in the same organization without inheriting one another’s access.

    This is a basic governance principle: access should follow responsibility. An agent should receive the capabilities required for its work without automatically inheriting access to every system its user or organization has connected.

    That makes the agent’s role easier to understand and limits the consequences of an incorrect request, an overly broad interpretation, or the accidental use of sensitive information outside its intended context.

    Collaboration Without Erasing Boundaries

    A Sentienta Workroom gives people and agents a shared place to work on a problem, project, or decision.

    Each agent can participate with its own instructions, context, enabled services, permitted operations, and approval requirements. Agents can respond to one another and contribute to the same discussion while their underlying service access remains separately governed.

    This distinction becomes especially important when an agent retrieves private information.

    Information retrieved from a user’s connected service is returned in a Private Workroom card. It is not automatically visible to other participants, other agents, or outside collaborators. The user can review the result and decide what, if anything, should be shared with the Workroom.

    The team can therefore share a relevant conclusion without exposing every contract, email, issue, file, or record that contributed to it.

    Consider a company deciding whether to make a significant product commitment to a customer.

    The Workroom includes agents representing sales, legal, engineering, and finance, along with the managers responsible for the decision. Sales can explain what the customer requested. Legal can clarify what the contract permits. Engineering can assess product readiness and implementation risk. Finance can describe the applicable budget constraint.

    Each contribution is informed by the systems suited to that role. The participants can compare their findings and reach a coordinated judgment without opening every underlying source to everyone in the Workroom.

    From Governed Access to Coordinated Action

    Connected Services do more than help agents retrieve information. They also allow the Workroom team to prepare and carry out follow-up work in the systems where that work belongs.

    In the customer commitment example, the engineering agent might review GitHub activity and related Jira issues. The legal agent might locate the applicable contract language in Drive. The sales agent might review customer correspondence and prepare a follow-up message. A project coordinator might prepare a meeting, while a communications agent prepares a summary for Slack.

    These are not necessarily the capabilities of one universal agent. Work is directed to the agent whose role and service permissions match the request. One agent may have several capabilities when its assignment requires them, but access is granted deliberately rather than assumed.

    The interaction remains natural. A user can ask Engineering whether a correction was deployed, ask Legal to find a contract provision, or ask the project coordinator to prepare a follow-up meeting. The appropriate agent uses its enabled services to retrieve information or prepare the requested action.

    Retrieval and analysis do not automatically authorize an external change.

    When a proposed action would send, post, create, update, or delete information, Sentienta can require the user to review what will change, which service and destination will be used, and which agent prepared the action.

    Approval applies to the proposed operation. It is not blanket authorization for unrelated future activity.

    The result is a layered governance path:

    1. The user authorizes a service connection.
    2. The user enables specific capabilities for specific agents.
    3. The user selects the agents participating in the Workroom.
    4. Each agent contributes from its permitted information sources.
    5. Retrieved private information remains private unless the user shares it.
    6. Consequential actions are handled according to the applicable authority and approval policy.
    7. The activity and outcome remain part of the operational record.

    Each layer answers a different governance question. Together, they allow a team to move from distributed evidence to a coordinated decision and then to accountable action.

    Governance Makes Greater Capability Possible

    Connected Services provide immediate practical value. They let Sentienta agents work with documents, communications, schedules, repositories, issues, and local resources without requiring users to carry information manually in and out of a chat window.

    Their larger importance, however, is how they prepare for a world of specialist agents.

    As departments develop agents around their own responsibilities and knowledge, those agents will need to collaborate across organizational boundaries. Collaboration cannot mean giving every agent access to every system. It must preserve meaningful distinctions among roles, information, and authority.

    Sentienta Workrooms provide the shared environment for that collaboration. Connected Services give individual agents bounded access to the knowledge and tools their work requires. Private cards support selective disclosure. Approval controls govern consequential actions.

    Governance is therefore not a restriction added after an agent has been made powerful. It is the structure that allows organizations to give agents meaningful capabilities safely.

    The future is not one agent with all of the company’s knowledge. It is a governed team of people and agents, each contributing the knowledge and capabilities appropriate to its responsibility.

  • How Sentienta Governs Agent Actions

    From governance principle to governance path

    Sentienta Governance

    A few weeks ago, we argued that enterprise AI needs a governance layer.

    That does not mean slowing adoption with more process. It means recognizing that AI is moving beyond individual experimentation. Employees are no longer using models only to summarize, draft, and analyze. They are beginning to work with agents that retrieve information, use tools, trigger workflows, post messages, create issues, and interact with the systems where company work happens.

    Once agents can act, governance becomes an operational question.

    It is no longer enough to ask whether a model produced a good answer. Organizations must also ask:

    • Was the agent authorized to use the service?
    • Was the requested action permitted?
    • Did it require human approval?
    • Which identity and credentials were used?
    • What happened when the action ran?
    • What record was preserved afterward?

    Over the past several weeks, we have introduced the capabilities announced with Sentienta V2, including collaborative Workrooms, governed agents, and controlled execution.

    This post looks at how those pieces work together. Agent work begins in a permissioned Workroom, moves through approved capabilities and controlled connections, pauses for human review when required, and leaves an activity record behind.

    The principle is straightforward: governance should not exist only in policy documents, administrative settings, or after-the-fact reviews. It should be built into how agent work is requested, authorized, executed, and recorded.

    That is what allows agents to move beyond conversation and participate safely in real organizational work.

    How Sentienta puts governance into the path of agent work

    Sentienta treats governance as part of the workflow—not as a policy document applied after the fact. Controls appear where work is requested, delegated, authorized, executed, and recorded.

    1. Workrooms establish the governed context

    A Workroom is more than a shared conversation. It provides the operating context for collaboration among people and agents: who is participating, which agents are present, who initiated a request, and who is permitted to manage or stop the work.

    Workroom permissions control who can join, contribute, invite participants, manage agents, and halt an active run. These controls establish the context used by Sentienta’s backend when it evaluates an agent request.

    Importantly, sharing a Workroom does not share a participant’s credentials, files, device access, or private services. Collaboration and authorization remain separate.

    2. Enterprise Agents turn individual experiments into organizational capabilities

    Many organizations begin using AI through individual experimentation. Employees create assistants, connect tools, and develop useful workflows—but the resulting capabilities may be difficult for the organization to understand, manage, or reproduce.

    Enterprise Agents provide a different model. They are organizational capabilities that can be made available to authorized users and Workrooms. Managers and teams can use approved agents without having to build and configure their own versions.

    This allows organizations to decide:

    • which Enterprise Agents are available
    • who may use and administer them
    • which services they may access
    • which execution environments they may use
    • and how their work is reviewed and recorded

    Enterprise Agents therefore turn agent capability from an individual configuration into a managed organizational resource.

    3. Roles and permissions separate use from administration

    Governance also requires control over who can configure the system. Sentienta distinguishes organization owners, administrators, and members.

    Owners and administrators manage organizational membership, Enterprise Agents, available services, and service policies. Members can use the capabilities made available to them without automatically gaining the ability to modify those capabilities or their underlying connections.

    This separation allows an organization to delegate useful agent capabilities broadly while keeping configuration and administrative authority appropriately limited.

    4. Brokered services mediate access to external systems

    When an agent needs to reach an external tool or organization-controlled system, Sentienta can place a governed service layer between the agent’s request and execution.

    The important question is not simply whether an agent can connect to a tool. It is whether the organization can control:

    • which services are available
    • which roles and agents may use them
    • which operations are permitted
    • where credentials remain
    • whether an action requires approval
    • and what activity is recorded afterward

    Permissions can distinguish among operations such as reading information, creating something new, updating an existing resource, sending a message, or deleting data. This avoids treating “access to a service” as a single all-or-nothing permission.

    5. The Enterprise Bridge creates a controlled execution boundary

    Some agent work must run in an environment controlled by the organization, for example, when it involves a private network, managed automation host, customer-held credential, internal service, or organization-operated connector.

    The Enterprise Bridge provides that execution boundary. It can advertise approved capabilities, receive authorized work, execute through controlled service connections, and return results to the Workroom.

    Credentials remain in the appropriate environment rather than being placed in an agent prompt or exposed to Workroom participants. If the required Bridge or service is unavailable, Sentienta can fail the request safely instead of silently choosing another user’s device, credentials, or connection.

    This creates a practical middle ground between powerless chatbots and unmanaged automation: agents can perform useful work, but execution follows a controlled organizational path.

    6. Approval gates preserve human control

    Not every agent contribution needs approval. Research, analysis, comparison, summarization, and drafting should remain fluid.

    Consequential actions are different. When an agent is preparing to change external state, such as posting information, creating a record, modifying a resource, or triggering another process, the applicable policy can require human approval before execution.

    The agent can prepare the proposed action and present it for review. An authorized person can approve or reject it before the external change occurs. This supports meaningful delegation without requiring the organization to accept unrestricted automation.

    7. Activity records make agent work reviewable

    A transcript shows what people and agents said, but governance also requires visibility into what the system did.

    Sentienta’s Activity records provide structured operational information about a run: the request, participating agents, important execution stages, service activity, approval events, outcomes, and failures. Enterprise governance records provide additional visibility into governed service actions and approval decisions.

    These records help managers understand how work progressed and investigate problems without presenting the transcript itself as a complete forensic audit.

    What this looks like in practice: posting a Slack update

    Imagine a product team using a Sentienta Workroom to coordinate a customer rollout. After resolving the open questions and agreeing on next steps, the manager wants to send a concise update to an internal Slack channel.

    An Enterprise Agent can summarize the discussion and draft the message. The governance question is what must happen before that message reaches Slack.

    1. The request begins in a Workroom

    The request starts where the team is already collaborating. The Workroom provides important context: who made the request, which Enterprise Agent is participating, what discussion led to the update, and who is permitted to manage the work.

    Workroom access does not automatically grant Slack access. It establishes the collaboration context that Sentienta carries into the authorization process.

    2. The Enterprise Agent prepares the action

    The manager asks the Enterprise Agent to draft an update. The agent reviews the relevant discussion, identifies the decisions and next steps, and prepares a proposed Slack message.

    Nothing has been posted yet. Drafting creates content inside the Workroom. Posting changes an external system, so it can be governed separately.

    3. Sentienta checks the service policy

    Before execution, Sentienta checks whether the organization has made Slack available through its governed service configuration.

    The system can evaluate whether:

    • the requester is an active organization member
    • the Enterprise Agent is allowed to use the service
    • the requester’s role permits the operation
    • the requested Slack channel is configured and allowed
    • and posting requires human approval

    The existence of a Slack connection is not enough. The specific agent, user, action, and destination must be permitted.

    4. An authorized person approves the post

    If policy requires approval, Sentienta pauses the action and presents the proposed message for review.

    The manager can confirm the content and destination, then approve or reject the request. If changes are needed, the request can be rejected and returned for revision. The external action occurs only after the required approval has been recorded.

    5. The Enterprise Bridge executes the approved action

    Once approved, the request is sent through the Enterprise Bridge. The Bridge uses the organization’s governed Slack connection, posts the approved message, and returns the result.

    Slack credentials remain in the controlled execution environment. They are not placed in the agent prompt, exposed in the Workroom transcript, or shared with participants.

    If the Bridge, service connection, or permitted destination is unavailable, the request fails safely rather than switching to an unmanaged connection.

    6. Governance activity records the outcome

    Sentienta records the relevant operational events, including the requested action, service and destination, approval decision, execution outcome, and any failure.

    The Workroom preserves the discussion that produced the update. The governance activity shows how the proposed action moved from drafting through authorization and execution.

    In an unmanaged workflow, “an agent posts to Slack” is merely a convenience feature. In a governed workflow, the agent accelerates the task while the organization retains control over authorization, approval, credentials, execution, and review.

    Agent actions do not bypass governance. They travel through it.

    Conclusion: managing agents that act

    The next era of enterprise AI is not only about producing better answers. It is about agents that can retrieve information, post updates, create records, trigger workflows, and interact with the systems where work happens.

    That creates a new management challenge. Organizations need to know which agents are trusted, which services they can use, what actions they may take, when human approval is required, how credentials are protected, and what record remains afterward.

    The goal is not to slow agents down. It is to give them a safe and accountable path from request to action.

    Agents will become useful at organizational scale only when companies can trust what happens beyond the chat window. Sentienta makes that work governed, visible, and manageable while still allowing people and agents to move quickly.