“Several agents”: what does it change?

• Shipping faster?
• Handing off part of the work?
• Actually running AIs in parallel?

“Several agents” = 6 meanings.

1 / 6

Several sessions.

1 · Several sessions

By hand, in the terminal.

1 · Several sessions

Claude + Codex, side by side.

1 · Several sessions

Or in the desktop app.

1 · Several sessions

Isolating each track: worktrees.

$ git worktree add ../jobs-a -b feat-a
$ git worktree add ../jobs-b -b feat-b
$ export HTTP_PORT=18082 COMPOSE_PROJECT_NAME=jobs_a

Isolated Git ≠ isolated Docker: one port set per track, or they collide.

2 / 6

Its subagents.

2 · Subagents

One agent, several subagents.

> Spin up a subagent to audit
  security, another one for performance.
  Then summarise both for me.

> Regression on the homepage since yesterday.
  3 subagents: the recent diff, the failing
  tests, the logs. What is the likely cause?

2 · Subagents

A burst, with no script: ultracode.

> ultracode: audit every endpoint under
  src/Controller/ — the missing access checks
# the agent writes the script, fires ~15 subagents,
# live progress in /workflows

One keyword: the agent writes the orchestration script itself and fires the whole burst in parallel. docs · paid plan · ≠ ultrathink (think harder)

2 · Subagents

What is tunable.

> Big refactor of the Payment module: one subagent
  per subfolder, on sonnet. Keep opus for the summary.

“38 subagents” = a session total, not 38 at the same time.

code.claude.com/docs/en/sub-agents · learn.chatgpt.com/…/subagents

2 · Subagents

Subagent ≠ background task.

Ctrl+B backgrounds either one · /tasks = the shared panel.

3 / 6

Delegation.

3 · Delegation

One agent calls another.

> Implement this cache.
  In parallel, ask Codex
  to do it too — we will compare.

> Implement this cache.
  Once it is done, ask Codex
  for a review before I look at it.

4 / 6

Profiles.

4 · Profiles

Defining distinct agents.

---
name: code-reviewer
description: Reviews code for quality and best practices
tools: Read, Glob, Grep
model: sonnet
---
You are a code reviewer. Analyze the code and give feedback.

Not just Claude, and not just subagents: Codex has its own (.codex/agents/*.toml) — but there is no shared format. docs

5 / 6

Homemade orchestration.

5 · Homemade orchestration

Industrialising it: a skill.

---
name: cross-review
description: implement, then have another agent review it
---
1. Implement the requested task.
2. Ask Codex for an independent review.
3. Reconcile the divergences, summarise.

Cheap, portable — but the loop is still driven by the model.

5 · Homemade orchestration

First, the agent on the command line.

$ claude -p "Audit the security of src/Payment/"
$ codex exec "Audit the security of src/Payment/"

Non-interactive: the prompt goes in as an argument. That is the brick you script.

5 · Homemade orchestration

… then a script to orchestrate.

import subprocess, concurrent.futures as cf
PROMPT = "Audit this module, be critical."
AGENTS = [["claude","-p"], ["codex","exec"]]
run = lambda c: subprocess.run(c+[PROMPT], text=True, capture_output=True).stdout
with cf.ThreadPoolExecutor() as ex:
    for out in ex.map(run, AGENTS): print(out)

Deterministic, versionable, 0 token for the loop. Repeated and stable → the script wins. E.g. prompt-fleet.

6 / 6

The orchestrators.

6 · Orchestrators

When a board drives the agents.

Adjacent

Adjacent: SpecKit, BMAD.

Roles and sequencing — not parallel agents on independent tasks.

Comparison

Which meaning for which need.

More agents =
more slicing.

Not more hands. The right move for the right problem.