A better operating model for agent work
Finish agent work
with proof it worked.
Use native tools for routine work, choose model and effort deliberately, and select a focused workflow when it helps. For authorized PR delivery, one CE owner settles review and CI through merge and proof. Add fleet placement explicitly when the work needs another machine.
A shipping example
One request
through the yard.
The operator asks to fix a retry path and merge the change. Native tools handle the implementation, with a useful workflow selected when needed. CE owns PR publication, review settlement, and CI. Completion includes the authorized merge and a focused check of the merged result. See how models are chosen.
you
“Ship the retry fix and get it merged.”
implementation allocation Sol 6.1 · effort medium
✓ implementation and focused checks
✓ CE commit/push/PR workflow
✓ merge reachable from main
post-merge proofverified
Illustrative transcript. Actual usage is a chat request to your agent, not a shell command.
The operating practice
Four habits.
One accountable system.
Make “done” provable
Verify the requested result, including merge and post-merge proof when shipping is authorized.
02Let the fleet converge
Read exact drift across a four-host macOS/Linux reference mix, then move each host toward declared state.
03Promote through evidence
Declare a skill once, review the item, prove the canary, and let signed history carry it outward.
04Place work with intent
Choose model and effort deliberately, and verify the destination when remote placement is requested.
enrolled and anonymized
one readable surface
an easy path
requested result
The reference fleet these docs were verified against.
01Ship a change
Carry one request through merge and prove the result on the base branch.
02Harden review
Concentrate correctness, security, and maintainability where the risk lives.
03Keep machines current
Turn exact inventory and drift into a fleet that converges itself.
04Distribute a skill
Promote one declared skill through item review, canary evidence, and signed history.
05Declare desired state
Give operators a readable map and machines an evidence-backed convergence target.
06Work across harnesses
Preserve the requested outcome across an explicitly selected, supported handoff.
07Build iOS and Mac apps
Put native builds and UI tests on disposable VMs with repeatable evidence.
08Run work on another machine
Place work where the project, identity, and transport are ready to receive it.
09Administer remotely
Keep remote operations named, resumable, and bound to verified identity.
10Control model cost
Compare accepted outcomes and total assignment cost, including retries and repairs.
Prove the practice once
Start with one
real change.
Install Railyard, complete a useful change on the machine you already use, and add fleet coordination when you need it.
Start a first delivery