Agentify: Controlled multi-agent coding workflows

01 / Problem & ownership

Balancing automated assistance with maintainer control and auditability.

02 / How it works

Deterministic orchestration from authorized GitHub issue to draft PR delivery.

Agentify installs into an existing GitHub repository. Maintainers submit or approve issues containing acceptance criteria and explicit candidate paths.

A persistent orchestrator works with a read-only planner and repository specialists to develop an implementation plan. One builder receives source-write authority on a task branch. After implementation, trusted code captures the changes and runs the approved repository checks.

A separate read-only reviewer evaluates the plan, diff, evidence, and acceptance criteria. Successful work is published as an unmerged draft pull request. The maintainer retains the final merge decision.

03 / Engineering decisions

Role segregation, repository policy enforcement, and accepted-commit learning.

One source writer: Planning and review roles do not receive application-source write authority. The builder operates within an approved path set, and trusted checks reject scope violations.

Validation bound to repository policy: Validation commands and relevant repository files are recorded during installation. Policy drift disables issue intake until the configuration is reviewed again.

Knowledge refresh after acceptance: Repository knowledge is updated from an accepted default-branch commit rather than treating an unmerged proposal as established repository truth. This updates stored context, not model weights.

// Trusted Validation Runner Execution
import { spawn } from "node:child_process";

export async function runValidationPlan(
  repoPath: string,
  command: string,
  args: string[],
  env: NodeJS.ProcessEnv
): Promise<ValidationExecutionResult> {
  return new Promise((resolve) => {
    const child = spawn(command, args, {
      cwd: repoPath,
      env: sanitizeEnvironment(env),
      stdio: ["ignore", "pipe", "pipe"],
    });

    let stdout = "";
    let stderr = "";
    child.stdout.on("data", (d) => { stdout += d; });
    child.stderr.on("data", (d) => { stderr += d; });

    child.on("close", (exitCode) => {
      resolve({
        passed: exitCode === 0,
        exitCode: exitCode ?? -1,
        stdout,
        stderr,
      });
    });
  });
}

04 / Evidence & limitations

Current test boundaries and security model scope.

Agentify is an early public project. Its current evidence includes maintainer-controlled qualification, security tests, and exact-package tests. It does not yet establish independent production adoption.

The controls constrain tool use, paths, publication, and access to GitHub write credentials. Repository validation remains unsandboxed and is not network-isolated. Automated review adds a separate assessment, but does not replace human review.