Sierra: why the best customer-facing AI agents run on simple architectures

Sierra co-founder Zak Reno-Wedeen and LangChain CEO Harrison Chase discussed the nature of AI-agents that actually work in production on the Max Agency podcast. The conclusion turned out to be counterintuitive: the most effective systems for customer service are not the most complex.

The Complexity Trap

Most teams creating AI-agents take the intuitive path: they add more layers, more tools, more orchestration. At Sierra, they came to the opposite conclusion after working with large corporate clients. According to Reno-Wedeen, agents with simple architecture are more stable and predictable in production. Each additional layer of abstraction is a potential point of failure, complexity in debugging, and unexpected behavior on edge cases. In customer-facing AI systems, this is critical: an error in the middle of a long chain can ruin the entire user experience.

• Fewer components — easier to debug and monitor
• Simple pipelines give predictable latency
• Easier to meet compliance and audit requirements
• Faster iteration when business logic changes
• Lower entry barrier for new engineers on the team

What is "org chart shipping"

One of the main anti-patterns Reno-Wedeen describes as "org chart shipping" — when an agent's architecture mirrors the organizational structure of the team that built it. If a company has a complaints department, technical support department, and sales department — the AI system develops three separate agents divided along the same lines. The problem is that the customer doesn't think in terms of internal org structure. Their actual task may span multiple departments simultaneously, and the system starts providing fragmented or contradictory answers.

"The best agent is one that solves the client's problem, not one that reflects the company's internal structure," —

Zak Reno-Wedeen.

The alternative is to design the agent around customer scenarios, not around how the company is structured internally. This requires a deep understanding of the customer journey, but delivers a significantly more cohesive user experience.

Outcome-Based Pricing

Sierra operates on an outcome-based pricing model — clients pay for solved tasks, not for the number of tokens or API calls. According to Reno-Wedeen, this model fundamentally changes the incentives for the entire development team. When payment is tied to outcomes, the team is forced to honestly answer in advance: what counts as success? This disciplines both product thinking and technical architecture. The agent is optimized for actual user outcomes, not system usage metrics.

In practical terms, this means that before going to production, the team must clearly define: what constitutes a "resolved request," under what conditions the agent is considered successful, and how to measure it. Without this work, any architecture — whether simple or complex — will optimize in vain.

What This Means

The AI-agent industry is moving from experimentation to maturity. Winners in the corporate segment are not those who pile on maximum technical complexity, but those who precisely defined the client's task and chose the minimally sufficient architecture for it. Sierra is one of the first exemplary examples of what this looks like in practice.