From Intent to Merge: A Hands-On Talk on SDLC in the Agent Era - Joseph Katsioloudes
This talk introduces 'Entire,' a developer platform designed to rebuild the Software Development Life Cycle (SDLC) for the agentic era. The core problem addressed is the loss of context and decision-making history (the 'missing middle') when using powerful, but forgetful, AI agents. Entire solves this by creating a semantic layer that versions not just the code diff, but the entire session history, allowing developers to track intent, micro-decisions, and evaluations from initial prompt to final merge. Key features include 'Checkpoints' (binding sessions to commits), 'Trails' (a reimagined Pull Request), and 'Runners' and 'Gates' for programmable code review and automated merging.
Key takeaways
-
The Need for Context in Agentic SDLC
0:28
Traditional Git only tracks the final artifact (the diff). When using AI agents, the 'missing middle'—the context, micro-decisions, and rabbit holes navigated—is lost. This makes code review incomplete, as the intent behind the code is often unknown. (0:00, 0:28)
-
Entire's Semantic Layer Approach
0:38
Entire saves conversations and context within the repository, creating a semantic layer. This allows developers to build organizational knowledge and avoid repeating mistakes across different agents and sessions. (0:38)
-
Checkpoints and History Preservation
7:10
A 'Checkpoint' is a mechanism that binds the entire session (prompt, transcript, tool calls) to a commit, creating a hidden branch (`.checkpoints`). This ensures that the full history, not just the code change, is versioned. (4:30)
-
Trails: The Reimagined Pull Request
20:00
The platform introduces 'Trails,' which are a reimagination of the PR designed for AI. They maintain the full conversation history and allow for structured review, including tracking confidence levels and findings. (12:00)
-
Programmable Review with Runners and Gates
22:10
Review is enhanced by 'Runners' (JSON files configured by prompt) and 'Gates.' Runners execute in sandboxes to address specific findings (e.g., 'address slope'), and Gates enforce configurable approval policies, moving toward automated merging. (13:30)