Superconductor's Arjun Singh on Multiplayer Agentic Engineering

Arjun Singh, co-founder of Superconductor, shared key insights on 'Multiplayer Agentic Engineering' at the AI Engineer World's Fair, detailing how teams and their AI agents can collaborate more effectively. Singh, whose background includes robotics and computer vision, and a previous successful venture with GradeScope, emphasized the importance of focusing on the human element within agentic workflows.

The Superconductor Approach to Agentic Collaboration

Singh outlined six lessons learned from integrating AI agents into Superconductor's workflows. The first lesson is to be model and harness agnostic. He explained that the best models and harnesses can change weekly, and teams should not be disrupted by these shifts. Open-weight models, like GLM 5.2, are becoming increasingly capable and cost-effective, offering flexibility without workflow disruption. Singh also cautioned that the incentives of token vendors may not align with a team's goals, highlighting the need for control over model choices.

The second lesson is to turn every human interface into an agent/human interface. Moving beyond agents being trapped on individual laptops, Singh advocated for making them accessible through various platforms like Slack, Superconductor's own app, or GitHub. The goal is to maintain the same agent session across these different interfaces, ensuring continuity and context. This allows for collaborative workflows where team members can seamlessly switch between environments, such as starting a task in Slack and continuing it in a desktop application.

Building on this, the third lesson is to make agent work visible and collaborative across the team. Singh demonstrated how an application view can show all interactions with a specific ticket, including who has been involved. This visibility is critical, especially when non-technical team members initiate tasks. He also stressed the importance of artifacts, such as screenshots or videos, generated by agents to make their work universally accessible and understandable.

Leveraging External Signals and Cloud Environments

Singh’s fourth lesson is to turn every external signal into code that your team can quickly evaluate. He noted that while systems can connect various data sources like Slack, emails, and bug trackers, agents often struggle to know what to act on without human intervention. Superconductor's approach involves automatically ingesting and prioritizing these signals. A practical example was their 'meeting bot,' which can listen to meetings, create tickets from ideas, and even prompt for action, streamlining the process from idea to implementation.

The fifth lesson is to set up cloud dev environments so agents aren't trapped on individual machines. This addresses 'lid anxiety' and ensures agents can run continuously without being tied to a specific laptop. More importantly, it allows for better security by giving agents access only to the necessary resources, preventing them from accessing sensitive data on developer machines. This also empowers non-technical team members to trigger real work, as they don't need their own development environments set up.

Finally, Singh emphasized the importance of benchmarking agents on your own codebase. Public benchmarks may not reflect how agents perform on a team's specific codebase. By selecting representative pull requests and benchmarking different agents, teams can obtain quality-versus-cost and quality-versus-time breakdowns. This data informs decisions on which models to use for different tasks, ensuring optimal performance and cost efficiency without disrupting workflows.

Key Takeaways for Agentic Engineering

Singh concluded with three key recommendations for teams looking to implement multiplayer agentic engineering:

• Get your codebase and agents working in a sandbox environment to unlock new workflows and collaboration possibilities.
• Integrate agents into all relevant human interfaces to enable seamless work and reduce context switching.
• Benchmark agents on your specific codebase and become model agnostic to stay at the forefront of cost, speed, and quality.

Superconductor's approach, as detailed by Singh, aims to make AI agents a truly collaborative tool, enhancing team productivity and the speed of innovation.