The next major challenge in artificial intelligence may not be building a more intelligent model. It may be engineering systems that can use existing intelligence reliably across complex, long-running workflows.
For Swarnamalya Mohan, an AI engineer and researcher whose work spans large language models, agentic systems, inference optimization, and enterprise AI, the distinction is fundamental.
While improvements in foundation models have accelerated the development of autonomous agents, the engineering infrastructure required to make those agents dependable has not always advanced at the same pace.
An agent that answers a technical question correctly is fundamentally different from one expected to investigate an infrastructure incident, retrieve evidence from multiple systems, coordinate diagnostic tools, and determine its next action without losing track of the original objective.
Swarnamalya’s work examines this gap between isolated model intelligence and the reliability required of AI systems operating within enterprise environments.
Why More Capable Models Do Not Automatically Produce More Reliable Agents
One of the central limitations of agentic AI is the difference between model-level accuracy and system-level reliability.
Traditional language model evaluations frequently measure performance against predefined questions or datasets. Autonomous agents, however, operate through sequences of dependent decisions.
A single unsupported assumption, incorrect tool selection, or outdated retrieval result can propagate through an entire workflow.
Even if every individual step in a workflow has a 99% probability of succeeding, a hypothetical 50-step sequence would have only about a 60.5% probability of complete success if those outcomes were independent. Real-world dependencies make the problem more complicated.
This distinction has shaped Swarnamalya’s interest in evaluating multi-agent systems beyond conventional accuracy metrics.
“The challenge with agentic AI is no longer just getting the model to generate the right answer. It’s ensuring that every decision, tool interaction, and intermediate step is grounded in evidence. A system that produces the correct output through an unreliable process is not necessarily a system we can trust in production,” Swarnamalya explains.
Rather than assessing an agent solely on whether its final response appears correct, she emphasizes the importance of evaluating intermediate execution steps, tool-call correctness, evidence quality, task completion, and failure recovery.
These considerations become especially significant in enterprise environments, where intelligent systems increasingly interact with operational infrastructure and business-critical workflows.
Applying Agentic AI to Production Incident Debugging and Triage
One practical application of these principles emerges in Swarnamalya’s engineering work on agentic AI for production incident debugging and intelligent triage.
Modern enterprise applications operate across distributed architectures involving interconnected services, APIs, databases, cloud infrastructure, and asynchronous processing systems.
When a production failure occurs, engineers often need to correlate application logs, distributed traces, infrastructure metrics, alerts, and historical incident records to determine the underlying cause.
The challenge is not simply retrieving this information. It is understanding the relationship between different signals and distinguishing the primary failure from its downstream symptoms.
Conventional incident investigation frequently involves manually navigating monitoring dashboards, identifying anomalous patterns, examining service dependencies, and reviewing previous incidents.
Agentic AI creates an opportunity to transform this process from a largely manual investigation into a structured, tool-assisted diagnostic workflow.
Engineering an Evidence-Grounded Investigation Framework
The technical direction of Swarnamalya’s incident-triage work centers on moving beyond conventional alert summarization toward evidence-supported diagnostic reasoning.
Her approach separates the investigation into specialized functions responsible for telemetry analysis, incident-history retrieval, root-cause hypothesis generation, and evidence verification.
At the center of this framework is a stateful investigation orchestrator, responsible for coordinating diagnostic actions and preserving relevant context as the investigation progresses.
Rather than providing a language model with a large collection of logs and requesting an immediate root-cause explanation, the workflow decomposes investigation into smaller, verifiable steps.
An incident may initially provide only an alert, an affected service, and a timestamp. The system must determine which additional signals are useful, retrieve relevant information, and update its hypotheses as new evidence becomes available.
For example, elevated API latency could originate from inefficient application logic, downstream service degradation, database contention, or infrastructure resource exhaustion.
A diagnostic agent might identify several plausible explanations, but those hypotheses should not be treated as confirmed conclusions without supporting evidence.
The distinguishing engineering challenge is therefore not merely generating hypotheses. It is designing the mechanisms through which hypotheses are tested, revised, and associated with observable signals.
Swarnamalya’s engineering approach emphasizes coordinating diagnostic actions through stateful orchestration and maintaining a traceable relationship between retrieved operational evidence and the hypotheses generated during investigation.
This distinction is important because a system can generate a convincing explanation while following an unreliable investigative process.
By emphasizing evidence verification and traceable execution, the architecture is designed to reduce the risk of treating plausible language-model outputs as definitive operational findings.
Integrating Language Models With Enterprise Observability
Implementing an intelligent incident-triage system requires more than connecting a language model to a monitoring dashboard.
An agentic investigation must interact with heterogeneous telemetry systems, maintain temporal context, retrieve historical information, and coordinate multiple diagnostic operations.
Swarnamalya’s approach uses a tool-augmented architecture in which a central orchestrator coordinates specialized diagnostic functions through controlled interfaces.
These functions retrieve and analyze operational signals such as logs, metrics, distributed traces, service dependency information, incident histories, and runbooks.
Retrieval-augmented generation provides relevant historical context, while structured working memory preserves the investigation state, including tested hypotheses, supporting evidence, and unresolved questions.
The system must also distinguish evidence from interpretation.
A log entry showing a connection timeout is an observation. A statement that a particular downstream database caused the incident is a hypothesis requiring additional validation.
Treating both as equally authoritative introduces the risk of unsupported root-cause attribution.
The verification process evaluates diagnostic findings against available telemetry, identifies contradictory evidence, and supports human escalation when the information is insufficient.
This produces a different operational model from unrestricted autonomous troubleshooting.
Instead of allowing an agent to move directly from diagnosis to production intervention, the approach prioritizes explainable findings, suggested next steps, and engineering oversight.
The deeper engineering contribution lies not merely in the individual components but in their coordination to support a traceable, context-aware investigation process.
Measuring the Impact of Agent-Assisted Incident Investigation
A central question for agentic incident triage is whether the technology produces measurable improvements over traditional engineering workflows.
Useful operational measurements include mean time to triage, manual investigation effort, root-cause identification accuracy, and the proportion of diagnostic findings supported by relevant evidence.
Swarnamalya’s approach emphasizes evaluation across both operational efficiency and diagnostic correctness, rather than relying solely on whether an agent produces a plausible final answer.
Historical incident replay can be used to assess how an agent investigates previously resolved failures, while expert-reviewed root-cause labels provide a reference for evaluating diagnostic accuracy.
Trace-level analysis can additionally measure tool-call correctness, evidence attribution, end-to-end latency, and the system’s behavior when information is incomplete or contradictory.
This broader evaluation is particularly important because faster investigations are not necessarily better if diagnostic accuracy deteriorates.
The distinction between a successful prototype and a dependable enterprise implementation ultimately depends on measurable performance under representative operational conditions.
For Swarnamalya, production incident triage illustrates a broader technical principle: the value of intelligent automation depends not only on what a system can accomplish, but on whether its decisions can be verified, reproduced, and safely integrated into operational workflows.
Context Engineering as an Architectural Challenge
Another emerging limitation of autonomous systems is their ability to maintain useful context across extended execution.
Long-running agents repeatedly retrieve documents, invoke tools, generate intermediate results, and update working memory.
As these interactions accumulate, the information available to a model can become increasingly noisy, redundant, or inconsistent.
Simply extending the context window does not necessarily resolve the problem.
An agent may have access to thousands of tokens of historical information while still failing to identify the specific evidence required for its next decision.
Swarnamalya has examined these challenges through her technical work and presentations on context engineering for long-running agents.
“Giving an AI agent a larger context window doesn’t necessarily make it more intelligent or reliable. The real engineering challenge is deciding what information to preserve, what to retrieve, and when to update the agent’s understanding. Context management is becoming just as important as model architecture,” Swarnamalya observes.
The architectural question is not merely how much information a model can process, but how an agent should select, preserve, compress, and retrieve information throughout a task.
Structured working memory, persistent state representations, retrieval policies, and selective context compression offer different approaches to this challenge.
Each introduces trade-offs involving latency, information loss, and retrieval relevance.
In incident investigation, these limitations are particularly visible.
An agent must preserve the original incident timeline, distinguish previously tested hypotheses from unresolved questions, and incorporate newly retrieved telemetry without discarding important findings.
Poor context management can lead to repeated diagnostic steps, outdated conclusions, and an inability to reconcile contradictory signals.
For enterprise AI systems, context engineering is therefore becoming an architectural concern rather than a prompt-design exercise.
From Plausible Outputs to Evidence-Grounded Decisions
The reliability challenge becomes even more significant when artificial intelligence is applied to operational infrastructure.
Enterprise information rarely exists in a single format.
Logs, structured metrics, infrastructure configurations, architectural documentation, and operational reports may each describe different aspects of the same underlying problem.
Swarnamalya’s research on evidence-grounded multimodal large language model agents for data-center capacity readiness examines how intelligent systems can reason over heterogeneous operational information.
This research direction addresses a persistent limitation of generative AI: the ability to produce plausible explanations without sufficiently verifiable supporting evidence.
Evidence-grounded architectures seek to establish stronger relationships between model-generated conclusions and the source information used to produce them.
Such systems can combine multimodal representations, retrieval, source attribution, and structured reasoning to support more traceable decision-making.
However, grounding itself introduces technical challenges.
Retrieved information can be incomplete, contradictory, or outdated. Multiple sources may describe the same infrastructure state differently, and a citation does not automatically establish that the resulting conclusion is correct.
The engineering objective is therefore to evaluate not only whether evidence was retrieved, but whether it is relevant, sufficient, and correctly interpreted.
This principle connects Swarnamalya’s research on multimodal infrastructure reasoning with her work on agentic incident investigation.
Both domains require intelligent systems to distinguish observed evidence from generated interpretations.
As AI systems become more autonomous, that distinction becomes increasingly important.
Evaluating the Entire Agentic System
The growing complexity of multi-agent architectures raises another question: what exactly should be measured?
A multi-agent system may contain specialized components responsible for planning, retrieval, execution, verification, and reporting.
Although specialization can improve modularity, it also introduces additional boundaries where context may be lost or information misinterpreted.
Evaluating only the final output can therefore conceal important weaknesses in execution.
Task-success rates, tool-call precision, evidence attribution, diagnostic accuracy, latency, resource consumption, and failure-recovery performance provide complementary perspectives on system quality.
More advanced evaluation approaches can use trace-level inspection, perturbation testing, adversarial scenarios, and controlled failure injection to expose weaknesses that conventional benchmarks may overlook.
These questions connect closely to Swarnamalya’s interest in evaluating multi-agent systems beyond accuracy.
“We need to stop evaluating AI agents solely on their final responses. In a multi-agent system, success also depends on whether the right tools were used, whether evidence was interpreted correctly, and how the system recovered from failures. If we cannot explain or reproduce how an agent reached a decision, we are not ready to give it greater autonomy,” Swarnamalya says.
Production incident triage provides a practical example.
An agent may correctly identify the affected service while incorrectly attributing the underlying failure.
Alternatively, it may produce a reasonable diagnostic hypothesis without retrieving enough evidence to support the conclusion.
In both cases, the final answer alone provides an incomplete measure of reliability.
Evaluating the full investigation process—including telemetry retrieval, hypothesis refinement, tool use, and uncertainty handling—provides a more meaningful view of operational performance.
The implication is that autonomy should be treated as a measurable system property rather than assumed from the presence of sophisticated language models.
The Efficiency Constraint in Production AI
Reliability is not the only architectural limitation facing enterprise AI.
As agentic systems grow more sophisticated, their computational requirements can increase substantially.
A single request may trigger multiple language model calls, retrieval operations, reasoning passes, verification steps, and external tool interactions.
These operations introduce latency and infrastructure costs even when individual models perform efficiently.
Swarnamalya’s technical interests also extend to neural network optimization and efficient inference, including research involving TensorTrain-based approaches in PyTorch.
Low-rank tensor representations offer one method for reducing the parameter and memory requirements of certain neural network computations, although practical improvements depend on model architecture, compression choices, and deployment hardware.
This connects to a broader challenge in agentic AI: balancing model capability with computational efficiency.
Model quantization, optimized inference serving, caching, batching, and selective model routing can influence the cost of operating intelligent agents at scale.
For operational applications such as incident investigation, this trade-off is especially consequential.
Additional reasoning and verification can improve diagnostic quality, but excessive model calls may increase latency during time-sensitive incidents.
The engineering goal is therefore not simply to minimize inference cost, but to allocate computational resources where they provide meaningful improvements in reliability.
