Different Organizational Models
Teams around the table already reflect radically different and sometimes contradictory operating models. On one end, some double down on structured, spec-driven approaches, trying to formalize how agents consume context, generate code, and validate outputs. Reviews shift from code to architecture and intent. At the other extreme, some have rebuilt entirely around agents, removing code review and CI, abandoning DRY principles, and designing highly scoped systems where agents operate in isolation. Systems are deployed directly to production, and failures are handled through fast iteration and rollback. These are not variations, they are opposing views of what engineering should become.
Beyond Code-Centric Workflows
There was a strong sense that current GitHub workflows are becoming structurally misaligned with agent-driven development. PRs, reviews and CI were designed for humans producing incremental changes, not agents generating code continuously. The core loop itself starts to break. Some expect new layers to emerge on top or alongside GitHub, where versioning extends beyond code to include prompts, specs and even agent sessions. The shift moves away from code as the primary artifact toward context, evaluation and traceability.
CI: The Main Bottleneck?
For some, CI is becoming the central nervous system of agent-driven development, the only reliable way to evaluate continuous flow of code. For others, it is pure overhead, too slow, too costly, and fundamentally incompatible with agent velocity. The underlying issue is that CI was designed for human-paced iteration. Speed of feedback becomes critical, especially for agents relying on evaluation loops. Accelerating CI is not enough, it needs to be rethought as an integrated feedback system. For Thierry Abaléa from Shipfox, while implementations will change drastically, CI as a practice and culture is not going away.
Hiring Being Redefined
Hiring practices are rapidly evolving and increasingly uncomfortable. Some teams no longer evaluate code at all and focus purely on outputs, giving candidates objectives and assessing what they ship. Others create adversarial setups where candidates must recover from broken agent outputs, testing their ability to steer systems. More traditional approaches, like whiteboard-based system design interviews, still exist but now focus on architecture and decision-making.