Running coding agents on a task graph instead of one long chat
The idea A feature usually breaks down into slices like this: graph LR T1[1. Data model] --> T2[2. API endpoint] T1 --> T3[3. Background job] T2 --> T4[4. UI page] T3 --> T4 T4 --> V[Validate merged result] V --> R[Code reviewer + live tester] R --> F[Fixer] Tasks 2 and 3 don't depend on each other, so there is no reason for them to wait. Task 4 needs both. Once everything is merged, the result…
The article presents a novel approach to software development that involves breaking down complex tasks into smaller, independent components and executing them concurrently. The process begins with a product requirement document (PRD), which is used to generate a task graph. This graph consists of various stages, such as defining the data model, creating API endpoints, executing background jobs, and updating the user interface.
Each task within the graph is represented by a small JSON file, containing the task ID, content, and dependencies. The tasks are stored in separate directories, allowing them to run independently and in parallel. The execution process is managed by a command-line interface (CLI) tool called graph-execution, which supports multiple coding agents such as Claude Code, Codex, OpenCode, and Qwen Code. The user selects the desired agent using the --provider flag.
The first step in the process is to write the spec, which is then used to generate the task folder and a test plan. Independent tasks are executed simultaneously, each in its own git worktree. Once a task is completed, it merges into an integration branch, and subsequent tasks that depend on it are allowed to run. The concurrency is set to three by default, but can be adjusted as needed.
After all tasks have been merged, a final validation check is performed. The merged branch is then reviewed by a code reviewer and a live tester, who both have access to the full spec, the test plan, and the changed files. The reviewers and tester work in parallel, providing feedback and any necessary fixes. If there are major issues, a fixer agent is triggered to make the necessary changes, and the validation process is repeated up to three times. Minor issues, however, are simply noted in the summary and do not trigger a fix.
The entire process is highly parallelized, with tasks running concurrently and independently whenever possible. This approach allows for faster development cycles, as developers can focus on individual components without waiting for other tasks to complete. The graph-execution tool handles all the heavy lifting, including creating pull requests, initiating human reviews, and managing the overall workflow.
To use graph-execution, the user must have Node.js 20+ installed, along with Git and a logged-in provider CLI for the chosen coding agent. The CLI is then run with the --spec flag, specifying the PRD file, and the --provider flag, indicating the desired coding agent. The web app can be accessed via http://127.0.0.1:7420, where the user can interact with the task graph and monitor the progress of the run.
Overall, this innovative approach to software development aims to streamline the process, reduce manual intervention, and improve the efficiency of complex feature implementations.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
