openclaw is the perfect agent orchestrator, but you're probably holding it wrong

current state of orchestrators

it seems that everyone on twitter is building their own agent orchestrator right now.

github, codex, and claude cloud coding are lacking in innovation and not great for moving fast -- so i've been very excited to try some of the new orchestrators.

orchestrators like ami, cmux, jean, and a bunch of side projects seem to be innovating on concepts like tickets, background agents, parallelism, notifications, etc. hell, i even have 2 of my own orchestrators i've been messing around with for the past 3 weeks (one as a jj-native github replacement and the other focused on my tmux workflow).

but the best one i've tried yet was not one i expected to be an agent orchestrator: openclaw (aka moltbot) (aka clawdbot) (aka stophie's stellar reboot into tech totally for fun)

enter the super-agent: openclaw

what makes openclaw a great orchestrator is its ability to handoff tasks to background sessions then notify you when it's finished.

this gives it a very simple conversational interface where you can do something like:

user: a user is having a bug in project X <screenshot>
openclaw: i'll start an agent to fix that
<few minutes later>
openclaw: the bug is now fixed
user: make sure there are tests to cover this
openclaw: sends new message to background agent
<few minutes later>
openclaw: 8 tests written and passing
user: push a pr
<a few seconds later>
openclaw: here is your pr link

when you put this together with the openclaw gateways (i.e. support for discord, slack, signal, etc), it makes it an incredibly smooth experience switching from your desktop to your phone.

my workflows

some misc behavior i've found works really well with openclaw:

git worktrees: tell openclaw to always do work in worktrees so you can keep programming in the main folder. consider giving it access to worktrunk too.
audio messages & walks: i now commonly go on long walks and iterate with openclaw when working on non-critical tasks or experimental projects (not production code 💀). openclaw knows how to pull the whisper model to run this 100% locally, though the api is faster.
tailscale: use tailscale for accessing dev servers & the openclaw dashboard
give it your repo paths: tell openclaw to update agents.md with paths to common repos so it doesn't have to search your computer every time
linear cli: give it access to your issues, let it rip
browser-agent cli: the fastest browser for agents on the market
daylight computer: i work pretty much morning to evening with a break for the gym every day, so i've been working on ways of improving my sleep. openclaw gave me an excuse to acquire a daylight dc-1 so i can now do code reviews & work with openclaw while winding down in the evening.

but you're probably holding it wrong

after a few days of using openclaw like this, i noticed it was unable to solve some really basic problems that claude code should have been able to do without issue -- even when i told it to use opus 4.5 or gpt 5.2 high. even weirder, i kept getting errors that the subagent ran out of context -- this never happens with claude code or codex.

something clicked and i realized -- openclaw is trying to act as a coding agent spawning instead of using claude code/etc.

openclaw has a powerful sessions_spawn tool to orchestrate background sessions. however, the default behavior is to use its own native sessions for coding.

this means that you're missing out on all of the benefits that the claude code/codex/etc harnesses provide.

coding harnesses matter

the reality is that in 2026, the harness matters almost as much as the model. without a good harness, you're missing out on:

better context management
better compaction
subagents (exploration and larger tasks)
fine-tuned system prompt

("harnesses" is just a fancy way of saying a "coding agent cli" instead of using models directly)

down the rabbit hole: the system prompt & skills

so i cloned openclaw to poke around how it's prompted.

here's the gist:

src/agents/system-prompt.ts
- snippet: If a task is more complex or takes longer, spawn a sub-agent. It will do the work for you and ping you when it's done. You can always check up on it.
skills/coding-agent/SKILL.md
- description: Run Codex CLI, Claude Code, OpenCode, or Pi Coding Agent via background process for programmatic control.
- read about skill descriptions if not familiar
- your query must match this description for the agent to load & use the skill for coding agents
- this skill also includes instructions on how to work with worktrees
- bonus: this skill auto-loads if it detects the coding agents in your $PATH, very cool
docs/tools/subagents.md & docs/concepts/multi-agent.md
- these need to be read via read tool
- likely will never make it into your context, since it needs to explicitly search for relevant terms

this means that:

the system prompt uses openclaw's session_spawn tool for native openclaw subagents, instead of coding agents
it has the ability to orchestrate coding clis, but needs to be explicitly told to use the skill

how to fix it

to fix it, just give openclaw this prompt:

Update your AGENTS.md to add this section:

Background Coding Work:

When I ask you to do coding tasks in the background or autonomously:
1. Use the coding-agent skill
2. Prefer Claude Code with Opus 4.5 model
3. Do NOT use sessions_spawn/sub-agents for coding work
For quick in-context edits, you can still code directly.

now test it with a task like:

create a new tanstack start project that says hello world

similarities with planner/worker swarm architectures

the cool part about this is this makes openclaw able to act as a planner orchestrating multiple tasks together, similar to patterns seen in cursor's long-running coding agents or most other successfully long-running coding agents.

current experiment: switching between sync and async coding

this is great for background tasks, but currently it's a bit clunky to switch from background agents to iterating with the agent in the foreground.

i'm currently experimenting with having openclaw automatically create a tmux session with nvim + cc open in the session, like this:

tmux new-session -d -s coding -c /tmp/my-task \; \
	send-keys 'nvim .' Enter \; \
	split-window -h -c /tmp/my-task \; \
	send-keys 'claude' Enter \; \
	select-pane -L \; \
	attach

this allows me to switch tmux sessions to jump into synchronous programming

bonus: `openclaw gateway wake` as your notification center

if you're not already using this command regularly, you should:

openclaw gateway wake --text "Done: [summary]" --mode now

this command will have openclaw wake up, process a command, and send you a message.

i've switched openclaw to being my main "notification center." anything i do in the background now uses this command.

a note on security

please:

sandbox your openclaw
do not give it access to main on any repos
do not give it access to any code that you do not want on the open internet
assume all of your tokens you give it can and will be exfiltrated (including github, anthropic, and openai)

i'm very careful about what i give openclaw access to, i find a new attack vector on this thing every few minutes.

have fun and be careful out there