Summary of Agent orchestration is so two-years ago

Overview of Agent orchestration is so two-years ago

In this episode of the Stack Overflow Podcast, host Ryan Donovan speaks with Sahil Jain, CTO of You.com, about how AI agent building has changed as models have become more capable. Jain argues that much of the industry is still over-investing in 2024-style orchestration, prompt engineering, and heavy harnesses, when newer models can increasingly handle planning and long-horizon tasks themselves. The real competitive advantage, he says, is shifting toward better tools, better data, better evaluation, and better product design around model capabilities.

Core Thesis: Agent Orchestration Is Losing Its Edge

The big shift

• In 2024, a lot of AI product value came from:
  • elaborate orchestration layers
  • prompt engineering
  • guardrails and scaffolding around models
• In the newer model landscape, many of those responsibilities are being absorbed into the model itself.
• Jain’s argument: the model is getting good enough that too much orchestration can become a liability rather than an advantage.

What still matters

• The orchestration layer, or “harness,” is still useful today.
• But its importance is decreasing over time as models improve at:
  • long-horizon reasoning
  • planning
  • multi-step task execution
• Builders should design products that improve as models improve, rather than relying on brittle logic that will age poorly.

Building AI Search for Agents, Not Just Humans

Search is changing

• You.com started as a consumer search engine, but Jain says the future user of search is increasingly AI agents, not just humans.
• Search is becoming infrastructure for models:
  • retrieving external knowledge
  • grounding model outputs
  • enabling agents to act in the world

Search as a tool, not the whole system

• Jain frames AI systems as having:
  • a model/engine that reasons and plans
  • a harness that orchestrates execution
  • tools that supply knowledge or enable action
• Search is one of the key tools because it connects the model to the outside world.

What Makes a Good Tool for an AI Agent?

Two broad categories of tools

• Knowledge tools
  • retrieve information not in the model’s weights
  • include public and private data sources
• Action tools
  • let agents do something in the world
  • examples: buying something, interacting with software, controlling robotics

What “better” means

• It depends on the tool:
  • lower latency
  • higher accuracy
  • better task completion
  • more reliable action success
• For search and retrieval tools, “better” often means:
  • more accurate results
  • stronger grounding
  • lower latency when needed
  • better alignment with the end task

How to Measure AI Products and Agents

Evals are hard, and benchmarks can mislead

• Jain says benchmark results are often not statistically significant or reproducible.
• Two major sources of variance:
  • task difficulty
  • agent unpredictability

Measure both the tool and the end-to-end system

• You can’t evaluate search in isolation if the real product is a search-powered agent.
• You need to measure:
  • the tool’s direct performance
  • the final output experienced by the user
• This matters because better tool scores do not always translate into better product outcomes.

Public benchmarks vs customer reality

• Public benchmarks are useful for signaling and progress tracking.
• But the most important question is whether the system actually solves customer use cases.
• Jain emphasizes that product teams need a clear definition of “good” for their specific use case.

Why Statistical Rigor Matters

Not all benchmark gains are real

• The You.com AI evals team found many benchmark claims in the industry are weak or non-reproducible.
• Jain stresses the importance of:
  • confidence intervals
  • reproducibility
  • controlling for variance
  • separating benchmark difficulty from agent instability

Practical takeaway

• If a system is variable, teams should determine whether that variability comes from:
  • the benchmark itself
  • the model/agent
  • the toolchain
• The goal is to make the agent as predictable as possible while preserving flexibility.

The Future of AI Products: Pick the Right Lane

Not every company should build generic chat

• Jain argues generic horizontal chat products will become harder to defend.
• Stronger opportunities are in:
  • tools that augment models
  • vertical AI products
  • data and evaluation infrastructure
  • RL / training environments for model improvement

Why this matters

• Companies should choose a space where model improvement helps them, not where it makes them obsolete.
• Jain’s metaphor: pick the lane where the wave pushes you forward.

Will Agents Become the Main Interface?

Probably, but not everywhere

• Jain believes agents will increasingly mediate tasks such as:
  • search
  • summarization
  • knowledge retrieval
• But he does not think agents will fully replace interfaces.

Rich interfaces still matter

• For tasks like:
  • ordering food
  • ride sharing
  • booking travel
• Human-friendly interfaces remain important, with agents embedded inside them.
• The likely future is a hybrid:
  • agents behind the scenes
  • richer UI/UX on top

What This Means for Publishers and Data Owners

Attribution is still unresolved

• Jain says search-based systems should provide citations.
• Citations help with:
  • transparency
  • attribution
  • publisher visibility
• But citations alone may not be enough if users don’t click through.

Compensation and licensing may become more important

• You.com is exploring partnerships with publishers and data companies.
• Jain suggests a future where:
  • high-value data is licensed
  • data owners are compensated
  • proprietary datasets power better AI experiences

Key Takeaways

• Stop overbuilding orchestration just because that was necessary in earlier model generations.
• Focus on tools, not just prompts:
  • search
  • proprietary knowledge
  • action-taking capabilities
• Evaluate carefully:
  • measure tool quality
  • measure end-to-end results
  • use statistically rigorous methods
• Build for the model era you’re heading into, not the one you started in.
• Agents will be everywhere, but not as a universal replacement for interfaces.
• Attribution and data licensing will be central issues as AI systems increasingly summarize and surface third-party content.

Notable Insight

The central idea of the episode is that AI products should be built to benefit from smarter models over time, not depend on fragile orchestration that the models will eventually outgrow.