DXC OASIS: The harness for mission-critical IT operations

The first wave of enterprise AI was about the model: larger, faster, better. Enterprises have learned, sometimes painfully, that model performance alone does not determine business outcomes. The real challenge is operationalizing intelligence.

Models reason, generate and assist, but they do not govern. They do not understand an organization’s boundaries, its risk appetite, its institutional knowledge, its regulatory controls, or when a human should oversee a decision. These preferences and controls are managed by the layer above the model: the harness.

LangChain gives the enterprise an open, durable, governable agent framework and the tools to scale and improve it through use. DXC OASIS makes it the harness for mission-critical IT operations. It connects agents to the customer's estate and data, grounds them in DXC's operational tooling and knowledge, and runs them in parallel, within policy and under continuous assurance. Customers retain ownership of their workflows, their data, and their competitive advantage.

Sixty years of running mission-critical systems in production has taught DXC what to watch for. DXC OASIS sits above the tools an enterprise already runs, connecting, rather than replacing. This means every system, signal and decision, across all vendors, platforms and services, is available in one view.

Through DXC OASIS, institutional intelligence is encoded into agents that operate continuously. In practical terms, that means the most senior engineer's judgment is applied at machine speed, across the entire infrastructure, without rest.

DXC OASIS is not another AI product for IT operations. DXC OASIS is the harness for mission-critical IT operations: the layer that turns AI models into safe, governed and economically viable operators. It seeks to rewrite an operating model that has stood for decades. Instead of operating a queue, agents enable teams to address issues in real time through an on-demand governance stream. Trust and assurance of agent behavior then becomes the center of IT operations.

DXC OASIS is in production on the open-source LangGraph runtime today.

1.1 Why queues exist

Queues exist because human capacity is limited. Operational tiers exist because knowledge is specialized. Escalations are necessary because no one person can know everything. Work shifts are necessary because people cannot work continuously. Queueing theory explains the rest. When a team runs near full utilization, as lean operations teams do, small variations in load produce large and unpredictable increases in wait. Priorities P1 to P4 are a rationing scheme for capacity, so P1s get worked on first while P4s can sometimes age past thirty days, and "SLA at risk" dashboards measure how close the team is to breaking service promises to customers. Site-reliability practice and conventional automation improve the margins as a "better version of today", but they don’t change the fundamental condition that underpins the queue.

Agentic AI changes the equation only if safety, governance and control are built into the architecture. Without that layer AI is an assistant. With it, AI is an operator. That layer is what’s known as the harness.

1.2 What DXC OASIS provides for every agent

Within DXC OASIS, every agent inherits the same capabilities by default: orchestration, so agents coordinate and delegate through a shared framework; context, so they reason over enterprise knowledge rather than training data; governance and guardrails, so every action occurs within defined authority; human oversight at the decision points where risk requires it; and assurance across the lifecycle, so every action is observable, auditable and reversible. None of this is built on a per-agent basis. Developers, and increasingly the operational SMEs who build agents directly, supply domain expertise; the harness supplies everything else.

1.3 How the harness eliminates the traditional IT queue

The queue that remains is a queue of decisions. Humans do not leave the loop; they move to the approval point. In place of a task queue, what accumulates is a short list of human judgments, each arriving with the diagnosis, proposed action and evidence already assembled.

DXC has spent sixty years learning that the best operational systems put the right human in the right place at the right moment, with exactly the context they need to make an informed decision. The 'good friction' that remains is deliberate, placed where consequence, novelty or irreversibility warrant it.

The goal is not to remove humans from operations. Instead, it is to remove humans as the bottleneck: humans provide judgment, agents provide capacity, and the harness orchestrates the collaboration between both.

LangChain’s technology is foundational to how DXC OASIS structures agent orchestration, delegation, state, human-in-the-loop and continuous improvement. Six capabilities are central:

2.1 Durable, governed execution

Mission-critical environments cannot depend on opaque autonomous behavior. LangGraph gives DXC OASIS deterministic entry and exit points, explicit state transitions and checkpointed execution that survives restarts and can pause indefinitely while a human decides. Rather than fragile sessions, every run is a resumable thread by design.

2.2 Ambient, event-driven agents

Most agents wait for a prompt. An ambient agent is started by an event: an alert, a ticket, a proactive telemetry insight, a schedule. It runs in the background, many instances at once, and surfaces to a human only when a decision is needed, through an inbox of pending interrupts rather than a chat window. This is what "no human prompt" means in the mechanism table, and is key as DXC OASIS moves from ticket-initiated to event-initiated operations.

2.3 Middleware as policy in the loop

LangGraph middleware attaches deterministic code at fixed points in the agent loop. Because the hooks are graph nodes they appear in the trace, are versioned with the agent and cannot be bypassed at runtime.

2.4 The Deep Agents harness

Deep Agents package the primitives for long-running, complex work: a virtual filesystem for scratch work; memory and skills; sandboxed code execution; planning; sub-agents as tools with isolated context; and context management that keeps long runs within bounds. Human steering is a first-class primitive with four verbs: approve, edit, reject, respond. Deep Agents are already being deployed as the next generation of DXC OASIS agents.

2.5 Evaluation as the test harness for non-deterministic systems

Unit tests cannot validate reasoning that may take a different path on an identical input. LangSmith provides observability and evaluation so that every run leads to improved performance. To do this, LangSmith supports curated datasets, offline evaluation before release, online evaluation of live traffic and experiments across versions, all built on traces of every agent step. DXC OASIS emits traces for every action today.

2.6 The right model for the right workload

Enterprise environments do not present a single AI problem. They present many. The harness is model agnostic, enabling individual model selection for each subagent or task, including open-weight and specialized models, all under one governance and observability framework.

LangChain makes the harness customizable and self-improving with runtime semantics, middleware, Deep Agents primitives, evaluation, tracing, and the tools to develop continual improvement loops. DXC OASIS makes it operable and accountable through the persona model, the knowledge hierarchy, tenant isolation and credential separation, the production runtime, and interoperability. This way, agents built by DXC, customers, partners and vendors run under one framework, with accountability for every action in a customer estate. The first part of this system is engineering. The second part is an operating model.

Without context engineering, an agent acts on general knowledge. This may be fine for a generic Linux server, but it’s insufficient for managing one that belongs to a particular customer. Context engineering is what makes an agent's reasoning, and therefore its actions, specific to the environment it operates within.

DXC OASIS deliberately structures context as a knowledge hierarchy: system-specific knowledge, customer-specific knowledge, DXC operational knowledge, vendor documentation and, as a final safety net, the model's own general knowledge. Retrieval works through a closest-to-issue first approach, so the most authoritative and specific information is prioritized.

The outcome of this approach is greater precision, stronger explainability, and higher trust. Institutional knowledge is transformed from being reference material into operational infrastructure that is used dynamically as required. The knowledge hierarchy structures which information an agent should trust for a given task first.

The LangChain framework is used to manage how an agent is given the right information at the right moment. It lets DXC OASIS decide, in code, what the agent sees before each step, which skills it is briefed with, what it remembers from previous work, how a long investigation is kept within bounds, and how a specialist sub-agent is handed only what it needs for a given task.

Through this model, context becomes a managed part of agentic operations: it is defined once, applied to every agent, and is visible in the trace of every run. That consistency is also the first condition of trust and assurance. An agent briefed the same way every time can be held to the same standard every time.

Removing the human from the throughput path also removes the judgment that was applied, implicitly, to every action. Through an agentic system, that judgment is rebuilt as a property of the system, and the control plane of operations shifts from queue management to agent trust and assurance.

Anthropic's AI Fluency framework names four competencies for working with AI: Delegation, Description, Discernment and Diligence. It was written for people. In mission-critical operations the "person" delegating to an agent is a policy that runs ten thousand times a night, so each competency has to be a property of the system, not of the individual. DXC OASIS carries all four.

4.1 Build-time assurance

The first thing that needs to be decided before an agent reaches production is what actions it is permitted to take. Delegation is determined by a governed policy, not the preference of an agent. An agent's authority to act is set by the reversibility, impact and verifiability of the action and by the agent's track record. As evidence of successful actions accumulates, the authority granted to the system can widen. Together, the criteria place every action in one of four tiers: observe, propose, act with approval, or act within scope. Every persona starts read-only.

The policy holds up because it is enforced in code, not explained to the agent. Policies that are important are injected as middleware at runtime, not merely guidelines. Middleware is deterministic code in the LangGraph run that masks sensitive data before a tool call, checks each action against the persona's authority, and raises an interrupt when an action is above its tier, so the run checkpoints and waits for a human decision. Nothing executes in between. Every agent runs on vetted components: the models it may call, the tools it may use, the knowledge it may read. There is a single secure path to production.

4.2 Run-time assurance

Every decision, tool call and outcome is emitted as a trace with a unique execution identifier. If an agent changed a configuration, the trace shows what it considered and why. Because every run is checkpointed, intervention is managed through a harness primitive rather than an emergency procedure.

On the operations wall, novel signals of agent behavior replace ticket counts and SLA timers: latency against target, cost, confidence and how often a human had to step in. This is the Agentic Operations Center, the counterpart to the NOC. It watches agents rather than queues and is in operation today.

4.3 One framework, every agent

Every agent in DXC OASIS runs under the same governance, whether DXC, a customer, a partner or a vendor built it. In this architecture, credentials never reach the model, secrets are managed separately from reasoning, an agent receives only the permission a specific action requires and customer operational data always stays with the customer’s tenant boundary.

Governance in the age of agents is not a policy document reviewed once a quarter. It is a property of the runtime, inherited by every agent and visible in every trace.

4.4 Worked example: preventing risky commands

A Linux agent operates inside customer production, where one wrong command can take a system down. A Linux SME designed a classification of high-risk commands, a confirmation flow, recovery behavior and an audit trail in an afternoon. The agent implementation is middleware injected around the tool call. Here’s what happens: a deterministic classifier recognizes an irreversible, high-impact command and raises an interrupt. The human-in-the-loop approves, edits or rejects before anything executes. It is the four competencies in their smallest complete form: an expert's description, a policy in code, a pause for judgment, and a record of all three.

Each competency a fluent human would exercise is present in this system, but as code in the loop, versioned and traced, rather than as the good habits of whoever happens to be on shift. Every one of those runs leaves a trace, and the trace enables future improvements to the system. DXC OASIS supplies the policy, the authority and the accountability. LangChain supplies the framework and observability tooling to support 24/7 operations.

Running agents is now table stakes. Establishing a harness that allows agents to compound improvements over time is what separates enterprises that are running AI, from enterprises that are building a durable advantage with it.

Harrison Chase puts it directly: "In software, the code documents the app; in AI, the traces do." The code says what an agent may do. The trace says what it did, on this input, in this run. Every run is therefore evidence, and evidence is where agent improvement starts.¹ This presents an opportunity in the context of mission-critical IT operations for the harness to not only redefine the operating model that has stood for decades, but to build into it the mechanisms for continual improvement.

An agent running in DXC OASIS in month six is not the same agent that was deployed in month one. It has absorbed the production signals and it has been tested against a growing library of real successes and failures. Its evaluations reflect the judgment of the experts who reviewed its outputs, drawing on decades of expertise. The learnings from every incident make the agent more robust against the next one.

In traditional IT operations, when something goes wrong, a review may be held, a runbook may be updated and lessons are passed on. This tribal knowledge lives in people and spreads only as fast as they are able to share it. DXC OASIS changes this. When an agent gets something wrong, the fix is made once, in code, and every agent within the fleet inherits that benefit. In this model, human preference is not consulted occasionally but rather is encoded continuously by design.

LangSmith is what makes the loop possible. It holds every trace, so every agent action is observable. Through evals, LangSmith scores every trace as they arrive, catching failures for engineering review. It turns each failure the engineer marks into a permanent test, so the fix can be validated and monitored for regressions. One platform carries the evidence from the moment a run finishes to the moment the next version is cleared to ship, and that is what lets improvement keep pace with agents that never stop running.

When agents work fifty times faster than people, whatever stays sequential, a manual trace review or a slow evaluation pipeline, becomes the bottleneck.

One of the most important discussions in enterprise AI is understanding where durable business advantage accrues, and recognizing that to have lasting impact, organizations must retain ownership of their intelligence.² This does not mean owning all intelligence, and it does not mean owning every layer in the stack. Where intelligence is central to the core business, it means controlling the components that determine how that intelligence behaves, how it is managed, and how it improves over time.

This philosophy holds true both for DXC as an operator of mission-critical IT and for our clients running their own businesses. Many enterprises wholly entrust IT intelligence to DXC so that they may focus upon their core business. Others rightly wish to retain ownership over their intellectual property, data and future technology choices related to their IT operations.

DXC’s response is three-layered:

• Open Harness. DXC OASIS codifies sixty years of DXC operational experience into the workflows, orchestration logic, skills, tool use, knowledge, memory and governance, while supporting interoperability with third-party and client-owned systems over open-standards protocols. Customers retain ownership of the operational capabilities they build on top of the platform.
• Open Models. Preserving model optionality allows the right model to be chosen for each workload, including open-weight models where they fit and allowing quality, cost, latency and privacy requirements to be met as they evolve.
• Open Runtime. Client deployments may span public cloud, private cloud, dedicated and sovereign environments as a configuration change, not a rewrite. Open-standards telemetry provides functional and cost accountability, while agent traces show how each agent behaved, building trust and feeding the improvement loop that compounds value over time.

That flexibility is embedded throughout DXC OASIS and is underpinned by LangChain's technology and core principles.

Where a client wants AI built into its core business, that intelligence should be entirely theirs. LangChain's technology is available to them directly, and DXC's AI consulting practices can help them build it to the same reference architecture.

The objective is straightforward: customers should control their future, not inherit ours.

DXC OASIS reached general availability in March 2026 and is scaling toward 130 customer environments across regulated industries and multiple geographies. Agents are available across twelve operational personas today, spanning Linux, Windows, SAP, virtualization, backup and patching, AWS, Azure, and network operations.

Across 56 measured use cases, DXC OASIS has reduced resolution time by 95 percent on average, with diagnostic accuracy of 92 percent or better validated against senior engineer judgment. Representative measurements:

For decades the industry optimized the queue rather than removing it. Agentic AI has the opportunity to do so, but not with models alone. Models reason, generate and assist. What they cannot do by themselves is operate safely inside a mission-critical estate. That takes a harness: context, orchestration, delegation policy, guardrails, evaluation, auditability, self-improvement and human oversight, built to open standards, intentionally designed, and working as one system.

The goal is not fully autonomous operations. Instead, it is more autonomy, more precisely governed, with a harness that hands agents progressively more work as evidence, confidence and controls scale together, watched from an operations center that measures agent behavior and operational value.

LangChain builds the framework that makes such a harness possible. DXC built DXC OASIS with it, for the systems that cannot fail. As virtualization redefined infrastructure and cloud redefined consumption, harness engineering will redefine how IT operations are delivered. The model is one component. The harness is where trust, assurance and enterprise value live.

DXC OASIS is the harness for mission-critical IT operations.

Footnotes

¹ Sam Crowder, The agent improvement loop starts with a trace, LangChain, 2026.
langchain.com/blog/traces-start-agent-improvement-loop 
 

² Harrison Chase, Own Your Intelligence, LangChain, 2026.
langchain.com/blog/own-your-intelligence