The Fifth Elephant
The Fifth Elephant

The Fifth Elephant Pune edition

Call for Proposals

All submissions

Aviral Tuteja

@aviraltuteja

Structured Coordination Over Prompt Chaos : Multi-agent systems with DSPy

Submitted Jan 9, 2026

Indicate the track in which your submission fits: Track 1 AI in Software Development Life Cycle (SDLC) Type of submission: 30 mins talk

Description

As LLM applications grow more complex, single-agent prompting quickly hits a ceiling. Planning, reasoning, tool use, verification, and safety all compete inside one prompt, creating systems that are fragile, opaque, and expensive to scale.
This talk presents multi-agent systems (MAS) as a software architecture pattern, not an emergent AI trick. Using DSPy, we’ll explore how agents can be built as specialized, contract-driven modules with explicit roles and communication protocols, enabling coordination, fault isolation, and systematic improvement.
I am not focusing on agent personalities or autonomy here. The emphasis is entirely on structure, showing how separating responsibilities leads to systems that are far more reliable than any single model.

Key Takeaways

  1. A clear definition of what an agent actually is: a role-bound DSPy module with contracts, not a “persona”

  2. Why single-agent ReAct pipelines break down as task complexity and tool count grow

  3. How multi-agent systems improve reliability through explicit separation of responsibilities

  4. How DSPy’s signatures, modules, optimizers, and traces map directly to multi-agent coordination problems

  5. A practical architectural mindset for building MAS as maintainable software, not prompt experiments

High-Level Flow

  1. The single-agent ceiling, where and why it breaks
  2. Agents as software components : roles, contracts, boundaries
  3. From pipelines to coordinated systems, specialist agents and shared state
  4. DSPy as coordination infrastructure, not just prompting
  5. Failure modes & guardrails with what goes wrong in real MAS

What This Talk Is Not

  • Not a prompt-engineering tutorial

  • Not a simulation of human teams or AI “employees”

  • Not framework evangelism

It is a grounded discussion of engineering tradeoffs, failure modes, and design patterns for MAS.

Who Should Attend

Ideal for:
Engineers building multi-step LLM workflows
Platform teams responsible for reliability and cost control
Architects evaluating agent-based system designs

Expected background:
Familiarity with LLM APIs and tool calling
Comfort with modular software design
No prior DSPy experience required

Speaker Bio

Aviral Tuteja is a Software Development Engineer at Unravel Tech, where he works on building production grade multi agent AI systems using DSPy. He has hands on experience designing modular agent architectures with specialized agents for business intelligence use cases, focusing on reliability, coordination, and observability rather than prompt-heavy demos.

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