Most AI coding tools still make you act like the operator. You write a long prompt, watch the terminal, wait for the next blocker, then decide what to do next. Vellum changes the shape of that workflow. Your assistant can use the coding agents orchestration skill to spawn Claude Code or Codex as a sub-agent, track the session, steer it when needed, and bring the finished work back into your normal flow.
What you delegate
You do not need to tell the assistant how to run tmux, poll a shell, or babysit another agent. You ask for the outcome: fix a bug, implement a feature, investigate a failing test, update a docs page, or review a risky change in an isolated worktree. The assistant decides whether to keep the work in its own tools or delegate it to a coding agent through the orchestration skill.
Can you have a coding agent fix the flaky onboarding test and open a PR?
Please keep it isolated, run the relevant tests, and send me the PR link with a short summary when it is ready.How it works
The assistant uses the Agent Client Protocol to spawn an external coding agent, such as Claude Code or Codex, as its own subprocess. That agent gets a clear task and a working directory, then runs autonomously with its own file editing and terminal access. Results stream back into the assistant conversation, so the assistant can keep track of what happened without pretending the user has to manage the session by hand.
For normal tasks, the assistant can launch a single coding agent in the project. For riskier work, it can create a separate git worktree first so the agent is isolated from the user's in-progress changes. If the agent goes in the wrong direction, the assistant can steer or abort the session. If the work should continue after the first task finishes, the assistant can start a follow-up agent session with the next instruction.
Once the code is ready, the assistant can inspect the diff, run checks, summarize the changes, and hand back the result in Slack, GitHub, or wherever the team is working. The important part is that the coding agent is not the product surface. The assistant is the orchestrator that decides when to delegate, monitors the run, and turns the result into a finished workflow.
Why this pattern matters
Coding agents are powerful, but they are still usually used like interactive tools. Vellum lets them become part of a larger assistant workflow. Your assistant can choose the right agent, run it in the right place, manage the lifecycle, and translate the output into a concrete deliverable. The user stops supervising a terminal and starts delegating engineering outcomes.
The outcome
A multi-step coding task can move from request to ready PR with far less active supervision. The user gives the goal, the assistant coordinates the agent, and the final handoff includes the code change, test status, PR link, and a plain-English summary of what changed.
