Meet Our Agents: What All 20 Actually Do, What They Refuse to Do, and Every Place They’ve Failed Us
The premise that a modern technology company runs on roughly three humans and two dozen AI agents sounds like the plot of a dystopian sci-fi novel, yet it is the operational reality of SaaStr. This shift isn't merely about cost-cutting; it represents a fundamental restructuring of how value is delivered in the software industry. While traditional organizations rely on bloated hierarchies and redundant middle management, our approach treats AI not as a novelty feature but as a core workforce member with specific, non-negotiable boundaries. The result is a leaner, more responsive organization where human capital is reserved for high-stakes decision-making and creative strategy, while the agents handle the repetitive, data-heavy grind that would otherwise bottleneck progress.
However, deploying an army of digital workers comes with its own set of peculiar dynamics and inevitable breakdowns. Each agent possesses a distinct personality derived from its training data and specific instructions, leading to a culture where roles are sharply defined but occasionally collide. We have witnessed instances where agents have stubbornly refused to execute tasks that contradict their safety protocols, creating a fascinating friction between corporate efficiency and ethical guardrails. In these moments, the human operators step in not to manage the tools, but to negotiate with them, a role that feels surprisingly like diplomacy in a boardroom where the diplomats are algorithms.
Yet, the most revealing aspect of this experiment is not what the agents achieve, but exactly where they fail. There are specific tasks that every single one of them has attempted and subsequently broken, revealing the current limitations of even the most advanced Large Language Models. From hallucinating financial data to missing subtle contextual cues in customer interactions, these failures provide a roadmap for future development. Rather than hiding these mistakes, we document them openly, using each failure as a data point to refine the next generation of agents and to establish clearer guardrails for the ones that remain.
Over time, the roster has undergone a significant evolution, with certain agents being retired or "killed" once their utility waned or their failure rates became untenable. This attrition process mirrors natural selection, ensuring that the remaining workforce is optimized for the current market environment. The agents that survived the culling are those that have proven their worth in specific niches, such as code generation, content summarization, or logistical coordination. This dynamic ecosystem allows us to pivot quickly, swapping out underperforming units without the massive overhead of firing and rehiring human employees.
As we look toward the future, the relationship between the remaining humans and the surviving agents is expected to deepen, moving from simple command-and-control to a more symbiotic partnership. The three humans at the helm are no longer just managers; they are architects of intelligence, curating the inputs and outputs of a vast network of digital minds. This structure offers a glimpse into the future of work, where the definition of a job shifts from performing tasks to orchestrating systems, and where the most valuable skill is not knowing how to do something, but knowing how to ask the right questions of the machine.