# Multiplayer agentic engineering — Arjun Singh, Superconductor

## Executive summary

The talk outlines 'multiplayer agentic engineering,' focusing on how human teams and AI agents can collaborate effectively across diverse systems. Key recommendations include making workflows model-agnostic, integrating agents into every human interface (Slack, GitHub), ensuring work visibility via artifacts, and treating all external signals (emails, meetings) as code inputs. Crucially, the speaker emphasizes that these advanced agentic workflows must operate within isolated cloud environments to enforce least privilege and prevent data exfiltration.

## Key takeaways

- Model Agnosticism is Critical: Do not rely on a single LLM or harness, as the best model changes frequently. Utilizing open-weight models (like GLM 5.2) allows teams to stay in control of costs and maintain workflow continuity.
- Agent Interfaces Must Be Universal: Agents should not be trapped on a single laptop or platform (e.g., Slack). The agent session must maintain context when moving between relevant interfaces like Slack, the desktop app, and GitHub.
- External Signals Must Become Code: Treat all external signals—customer calls, meetings (e.g., a Google Meet bot), bug reports from Sentry, or emails—as inputs that can automatically trigger and prioritize work for the coding agent.
- Work Must Be Visible Everywhere: To improve collaboration, agents should make their work visible across all platforms using standardized artifacts (screenshots or videos), eliminating context switching.
- Use Isolated Cloud Environments: Running agents in a cloud sandbox is essential for security, enabling non-technical staff to trigger real work without having development environments on their local machines. This enforces the principle of least privilege.

## Technical details

- Agent Workflow Architecture: The ideal setup requires agents to operate in an isolated cloud environment, allowing seamless interaction from multiple interfaces (Slack, desktop app, GitHub) while maintaining a single, consistent agent session and context.
- Security & Least Privilege: To mitigate risks like 'lid anxiety' or unauthorized access, agents must be restricted by configurable network sandboxes. This prevents exfiltration of code/secrets and requires explicit permission for new access.
- Agent Benchmarking: Teams should benchmark agents using pull requests from their own codebase, rather than relying on public benchmarks (e.g., SWE bench is Python while the company uses Ruby on Rails). This provides accurate quality vs. cost/time data.

## Practical implications

- Implement a sandboxed cloud environment for all agentic workflows to enforce least privilege and secure data.
- Develop integration layers that allow agents to maintain context across multiple human interfaces (e.g., Slack, GitHub).
- Establish automated processes to ingest external signals (emails, meeting transcripts) and convert them into actionable tickets or code tasks.
- Conduct internal benchmarking of LLMs using proprietary pull requests to determine the optimal balance between cost, speed, and quality for specific engineering tasks.

## Topics

AI Agents, Software Engineering Workflow, LLM Architecture, DevOps, Cloud Security, Superconductor, GradeScope

Source: https://www.youtube.com/watch?v=OL7kfezynJM