All Insights
The New Rules of Building Agentic Software
Vista’s Agentic Factory has spent the past year helping portfolio companies design and launch agentic products. Across dozens of engagements, that work has challenged many of the basic assumptions behind traditional software development.
The changes are showing up in three areas in particular: how workflows are designed, how product teams are organized and how companies manage the economics of AI-powered products.
Here’s what we’ve seen emerge as a new set of guidelines for building agentic software that is reliable and economical at scale.
Supporting the Workflow → Running the Workflow
Traditional software supports the users who carry out a workflow. Agentic AI performs work autonomously, completing individual steps or even managing entire processes end to end. That puts greater demands on workflow design.
Teams should map workflows step by step, then determine how each step should be executed. Some require an agent’s ability to reason and make decisions. Others are better handled by deterministic software: fixed, rules-based logic that produces the same result every time.
From there, different points in the workflow call for different types of agents, each specialized for the task at hand. After dozens of Agentic Factory engagements, we’ve identified four agentic archetypes:
- Smart Sidekicks work directly alongside users, helping them complete everyday tasks. Because users initiate and interact with them directly, they are among the most intuitive agents for organizations to adopt.
- Fixers That Never Sleep, or always-on agents, operate continuously in the background, monitoring systems to identify and resolve issues, such as security threats, before they escalate.
- Workhorses handle high-volume workloads, processing large amounts of data across complex deployments.
- Conductors manage complex, multi-step workflows, orchestrating multiple agents to complete work with minimal human oversight.
Whatever the archetype, enterprise agents require deterministic execution around AI reasoning. AI provides judgment; traditional software provides consistency and control. Designing the right balance between the two can be critical to running enterprise workflows reliably.
Corporate Hierarchies → Cross-functional Pods
Agent-assisted engineering has made prototypes faster and less expensive to build, allowing teams to test products with customers earlier and iterate more often.
But as engineering gets faster, teams are contending with a new bottleneck: communication. The constraint is becoming how quickly decisions and feedback can move between people.
Software development as we knew it was sequential. Product teams scope ideas and secure approval from leadership, then hand requirements to engineering. Any changes have to travel back through that process for approval and execution. That structure becomes a bottleneck when development cycles shrink from weeks or months to days.
For a more agile approach, companies should move from siloed teams to cross-functional pods. They bring product, engineering and UX together with domain experts and leadership, allowing decisions that once moved sequentially across functions to happen in parallel.
Vista’s Agentic Factory implements pods throughout the development process, from initial scoping through the engineering build. With the relevant experts and decision makers working together, teams can complete several traditional sprints’ worth of work in a matter of days. All companies that have progressed through the Agentic Factory have since adopted pod models for future product builds.
Most Powerful Model → Most Efficient Model
Traditional software has long benefited from low marginal costs: build it once, sell it many times. Agentic software introduces a different cost structure.
Every action an agent takes consumes compute, and costs rise with the complexity of the task. Vista estimates that a simple chat query using Claude Sonnet 5 costs approximately $0.002, while an agentic workflow can cost $0.50 or more — a 250x increase.1
For agentic software, margin is largely a design decision.
The most powerful model is not necessarily the right model for every task. Teams should match model capability and cost to the level of reasoning each step actually requires. Two approaches can make a significant difference:
- Break workflows into smaller tasks. Discrete steps can often be handled by a few smaller, cheaper models coordinated together, rather than one expensive model.
- Default to the least powerful model that can reliably do the job. Open-weight and older models can perform at or near frontier levels on many enterprise tasks at a fraction of the cost. The Agentic Factory built an agent that tests a company’s prompts and evaluations against open-weight models to determine where they can be confidently substituted, typically reducing model costs by 50 to 60 percent.
Some tasks do require heavy reasoning. Those are often the tasks customers value most when automated, and pricing should reflect the higher cost of delivering them.
In every case, the goal is balance. Teams need to understand what each task costs, what level of reasoning it requires and what value it delivers, then design the product accordingly.
New Rules, New Results
Agentic software requires teams to rethink more than the technology itself. Workflows should balance AI reasoning with deterministic execution. Development teams need to make decisions as quickly as they can now build. And successful products must be designed around the cost of running them, not just the cost of creating them.
Getting those decisions right is what separates promising agentic prototypes from products that can operate reliably and economically at enterprise scale.
For teams that succeed, the upside is twofold: new revenue from products that perform valuable work, and stronger margins from delivering that work more efficiently.
Learn more about how Vista’s Agentic Factory is putting these rules into practice across the portfolio.
