Orchestrate parallel coding agents with Polyaxon
Run parallel coding agents in separate Polyaxon workspaces, collect candidate patches, evaluate independently, and merge only reviewed changes.
Parallel coding agents can explore independent changes without competing for the same working directory. The useful output is not a set of remote edits arriving directly in the main branch; it is a set of candidate patches with evidence.
Polyaxon sandbox services can provide a workspace for each task. An application-owned coordinator assigns work, collects results, and leaves promotion under the team's control.
Give each task its own source tree
Start every task from a recorded base commit. Create a separate service and workspace for each agent, using the same reviewed image and dependency configuration where the tasks require equivalent environments.
One agent might improve data-validation error messages while another investigates a preprocessing bug. They can share the starting revision without sharing mutable files.
The Git workflow guide covers repository operations through sandbox access. Configure private-source credentials narrowly; generated code does not need broad repository administration or release permissions.
Keep coordination in trusted code
The coordinator maps task identifiers to authorized projects and run UUIDs. It sets deadlines, limits parallel tasks, and records which input revision each agent received.
Use sandbox process and filesystem interfaces to execute commands and retrieve results. A model should not be able to redirect another task into an unrelated workspace.
Keep a cumulative budget for the whole request. Parallelism can reduce elapsed time while increasing total resource usage, so a per-service timeout alone is not enough.
Collect a candidate manifest
Require each task to return a structured record:
| Field | Why it matters |
|---|---|
| Base commit | Detects incompatible starting points |
| Patch or candidate commit | Identifies the proposed change |
| Changed-file list | Supports scope review |
| Execution run UUID | Connects the result to its workspace |
| Evaluation references | Shows which checks were performed |
| Known limitations | Prevents incomplete work from appearing finished |
A simple unstaged Git diff can miss new files or staged changes. Define a complete collection procedure that accounts for tracked, staged, and untracked content, and checks for accidental secrets before persistence.
Store approved candidates through Polyaxon artifacts or a scoped source-control workflow.
Evaluate outside the agent's workspace
Apply each candidate to a clean copy of the recorded base and run the team's approved checks in a separate job. Keep evaluator code and fixtures outside the candidate's writable environment.
Record correctness, scope, and relevant runtime or quality measurements. Use run comparisons when evaluating alternative solutions to the same task.
Two individually valid patches may conflict or change shared assumptions. Evaluate the combined result after integration rather than assuming independent success guarantees compatibility.
Merge deliberately and release the workspaces
Require review before applying a candidate to the main working tree or opening a release path. Keep merge and deployment credentials in the trusted promotion process.
Preserve useful patches and reports before stopping the services. Use termination controls to bound forgotten workspaces and track cleanup completion in the coordinator.
This is a general integration pattern using Polyaxon primitives, not a claim of a bundled coding-agent coordinator. It gives parallel agents room to work while keeping source changes, evaluation, and final acceptance inspectable.