The Fifth Elephant
The Fifth Elephant

Enterprise AI in Production: Mumbai Call for Proposals

Share what it takes to run AI inside an enterprise: architectures, trade-offs, and lessons from production.

All submissions
Swetha A

Swetha A

@swetha_03

When Good Agents Disagreed Until We Built Shared Understanding

Submitted Oct 2, 2026

Abstract

Multi-agent systems are often built around orchestration: an orchestrator routes requests to specialist agents, coordinates their work, and brings their outputs together. But in production, we encountered a more fundamental problem: every agent could be doing its job correctly, while the system as a whole was still wrong.

We saw this while building a production multi-agent assistant for enterprise planning. As real conversations grew more complex, ambiguity began to compound across agents and turns. The same entity could acquire different meanings, context could be interpreted differently by different agents, and locally correct decisions could lead to globally inconsistent outcomes.

The root cause wasn’t orchestration. It was semantic drift.

In this talk, I’ll share how we evolved the system around a shared semantic foundation using intent understanding, entity grounding, disambiguation, and explicit semantic state to establish a consistent interpretation before downstream agents act. We also separated semantic state from conversation history and introduced delta-based history management, allowing the system to carry forward what actually changed rather than repeatedly reconstructing meaning from an ever-growing conversation.

Through this evolution, we found that reliability came less from adding more agents or increasingly complex orchestration, and more from establishing clear boundaries between language understanding, semantic state, conversation history, memory management, deterministic logic, orchestration, and agent reasoning.

I’ll walk through the production failures that led us to these changes, the architectural trade-offs involved, and the practical patterns that helped us make the system more reliable under real-world usage.

The central lesson is simple: orchestration can coordinate agents, but it cannot resolve inconsistent meanings. Before agents can collaborate reliably, they need to share an understanding of what the user actually means.

Where appropriate, production examples will be anonymized to focus on the engineering patterns rather than the underlying product or business.

Architecture Summary

Link to diagram:
https://drive.google.com/file/d/12F5dVvEnNGbkeTylTBWrwV16g-FnvqEQ/view?usp=sharing

https://docs.google.com/presentation/d/1vyT2y5nRKuQlabmnNSt3FLAoPsX-ebg0iVa4S-jYtCk/edit?slide=id.g40a07e64c20_3_2#slide=id.g40a07e64c20_3_2

Failures observed initially

1. The same entity meant different things to different agents.
A term the user used once was resolved by one agent to one object and by another agent to a different one. Each agent’s output was correct for its own interpretation, but the combined answer contradicted itself.

2. Meaning changed mid-conversation as history grew.
In 30-40 turn conversations, agents re-read the full transcript on every turn and weighted old turns differently. A clarification made early could be lost or overridden later, so the same question got different answers at turn 5 and turn 25.

3. Ambiguity was silently resolved instead of surfaced.
When a request was ambiguous, downstream agents guessed and carried on. The error only appeared several steps later, in the final answer, which made it hard to trace.

4. Orchestration changes did not fix any of this.
Our first responses were to improve routing, add agents and refine prompts. None of these addressed the problem, because it was a disagreement about meaning, not a coordination problem.

Production Evidence after the fix

Scale: Approximately 350 users, 700 conversations per day, average turns per conversation is 30-40
Time in production: 1.5 years now
Results observed after making appropriate fixes: Drop in cross-agent inconsistencies, fewer clarification loops or user corrections, task success rate, token reduction from delta-based history, latency change
How it was measured: A labelled set of real conversations, before/after comparison, LLM-as-judge with human spot checks, manual analysis, Langfuse observability and tracing

Key takeaways

  • How semantic drift emerges across multi-turn conversations and multiple agents.
  • A practical pattern for introducing a shared semantic layer before downstream agent execution.
  • How intent, entities, grounding, disambiguation, and state transitions can work together in an agentic architecture.
  • How state management and conversation memory management can be designed explicitly rather than relying on raw conversation history.
  • Where to draw the boundary between LLM-driven reasoning and deterministic application logic.
  • Production lessons from evolving a multi-agent system beyond the initial “orchestrator + specialist agents” architecture.

Target Audience

  • AI/ML Engineers building production-grade agentic systems
  • Software Engineers and Developers working with LLM-based applications
  • Solution Architects and Technical Leads designing multi-agent architectures
  • Engineers working on conversational AI, enterprise assistants, and workflow automation

Anyone moving agentic systems from prototype to production and dealing with real-world failures

Preferred format:

30 minute talk

Speaker Bio

I’m a Solution Consultant at Sahaj Software. My work spans multi agentic systems and the practical applications of GenAI in engineering.
I enjoy exploring how AI technologies can augment human creativity and decision-making. My research has been presented at the International Conference on Data Analytics and Management, and I’ve spoken at multiple DevDays events and Fifth Elephant conferences, sharing insights on AI agents, AI-assisted software development, and emerging agentic architectures.

Linked In:

https://www.linkedin.com/in/swetha0302

Past talks & workshops:

While https://youtu.be/zWWWi5tfkn4?si=SXbpzCFN62GDcAxa was about what are agents and how to build them from scratch, this proposal will be to share the production failures, architectural trade-offs, and practical patterns that transformed our agentic system

https://docs.google.com/presentation/d/1vyT2y5nRKuQlabmnNSt3FLAoPsX-ebg0iVa4S-jYtCk/edit?slide=id.g40a07e64c20_3_2#slide=id.g40a07e64c20_3_2

All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures