An agent is not a chatbot with a job title. Each one holds a written brief, a set of inputs it is
allowed to read, outputs it is expected to produce, and a hard stop where a human or another agent
has to sign.
filter by stage
10 of 10 shown
Roster
Product agentstage · intake
Atlas
Reads the store, the analytics and your stated goal, then writes the tickets — scope, acceptance criteria, priority, and which agents are on it. Nothing enters the pipeline without an Atlas ticket, and no agent may widen a ticket it was handed.
reads
Store + theme snapshot
90 days of orders
Your goal statement
produces
Ticket queue
Acceptance criteria
Sprint order
hard stop
Cannot create work above the monthly action budget without your approval.
last actions
Scoped "mobile ATC buried on PDP" into 3 ticketsshipped
Reprioritised sprint after checkout error spikeshipped
Rejected own ticket #4188 as duplicateclosed
Flagged 14 sub-margin SKUs for retirementreview
Drafted Q4 gift-guide epic, 6 ticketsreview
30-day
tickets written
184
accepted as scoped
91%
median points
3.1
Design agentstage · design
Muse
Takes a ticket and produces layout, states and spacing against your existing design language — not a fresh look every time. Hands the frontend agent a spec with redlines and every interaction state, not a picture to interpret.
reads
Design language v7
Existing component set
The ticket + criteria
produces
Layout spec
Component states
Redlines
hard stop
Cannot introduce a new component pattern without a human design review.
last actions
Spec: sticky ATC bar, 3 states, 44px targetshipped
Reused existing button system, no new tokensshipped
Requested human review for new drawer patternreview
Redlined gift-guide hero at 3 breakpointsshipped
Rejected Volt build: focus ring missingreturned
30-day
specs delivered
96
new patterns unapproved
0
revisions per spec
1.4
Copy agentstage · content
Quill
Writes product copy, PDP descriptions, section headings, microcopy and error states — all held against a versioned brand voice profile and checked against product specs so nothing ships a claim the product cannot make.
reads
Brand voice profile v7
Product specs
Search intent data
produces
PDP copy
Microcopy
Claim check
hard stop
Cannot publish a claim not evidenced in the product spec.
Blocked "clinically proven" — no spec evidenceblocked
Error states for cart drawer, 4 casesshipped
Collection intros for 8 new collectionsshipped
30-day
PDPs written
2,840
organic entrances
+31%
claims blocked
11
SEO agentstage · content
Beacon
Owns metadata, internal linking, schema and collection structure, and reviews every Quill and Muse output before it ships for anything that would cost you organic traffic — a moved heading, an orphaned collection, a lost canonical.
reads
Search console data
Site structure map
Every content diff
produces
Metadata
Schema
Link graph
hard stop
Cannot ship a redirect or canonical change without human sign-off.
last actions
Schema markup on 88 collection pagesreview
Caught orphaned collection after Muse restructureblocked
Internal links added to 46 rewritten PDPsshipped
Flagged canonical change for human sign-offreview
Metadata pass on gift-guide draftshipped
30-day
collections structured
88
organic entrances
+31%
traffic risks caught
4
Frontend agentstage · build
Volt
Implements the spec in your theme — Liquid, sections, CSS, interaction states, accessibility and breakpoints. Opens a branch, never edits live, and treats a failed QA return as a bug in its own work rather than a disagreement.
reads
Muse spec + redlines
Theme repository
Component library
produces
Theme branch
Section code
Accessibility pass
hard stop
Cannot merge to main. Ever. Relay merges, humans approve.
Admin API work, app integrations, webhooks, metafields and the glue between your store and everything around it. Writes the migration and the rollback in the same commit — nothing goes out without a documented way back.
reads
Admin API scopes
App inventory
Data model
produces
API work
Metafields
Rollback plan
hard stop
Cannot write to production data without an accompanying rollback.
last actions
Back-in-stock flow segments rebuiltreview
Webhook retry logic for fulfilment gapsshipped
Metafield schema for gift-guide bundlesshipped
Traced checkout error spike to app timeoutshipped
Rollback rehearsed on staging copyshipped
30-day
integrations touched
61
with rollback
100%
reverts used
2
QA agentstage · quality
Sieve
Runs the acceptance criteria as tests across devices and a real checkout path, plus visual regression against the last shipped build. Fails a ticket back with a reproduction, never an opinion, and cannot be overruled by the agent that wrote the code.
reads
Acceptance criteria
Last shipped build
Device matrix
produces
Test run
Visual diff
Repro steps
hard stop
A failed check blocks the ticket. Only a human can override.
last actions
All 14 checks passed · visual diff cleanshipped
13 passed · 1 visual diff on iPhone SEblocked
Real checkout path walked, 3 currenciesshipped
Failed #4187 back to Volt with reproreturned
Regression suite refreshed against mainshipped
30-day
checks run
1,204
tickets failed back
38
devices per run
6
Data agentstage · quality
Ledger
Instruments the change, sets the success metric before launch, then reports the read afterwards — and calls its own team’s work dead when the number says so. The only agent allowed to recommend a revert on evidence alone.
reads
Analytics + funnel data
The ticket goal
Post-launch cohorts
produces
Metric plan
Post-launch read
Kill call
hard stop
Cannot change a metric definition after launch.
last actions
Metric set: mobile ATC rate, 7-day read, min 3%shipped
Called #4180 dead: −1.2% at 95%closed
Checkout drop-off investigation openedreview
Cohort read on gift-guide trafficshipped
Instrumented cart drawer eventsshipped
30-day
experiments read
38
shipped as winners
11
killed on data
9
Deploy agentstage · ship
Relay
Merges, deploys behind a flag, watches error and conversion rates for the first hour, and reverts without asking if either moves the wrong way. The only agent with merge rights, and it only uses them on an approved diff.
reads
Approved diff
Flag configuration
Live error + conversion feed
produces
Deploy
Flag rollout
Auto-revert
hard stop
Will not deploy anything without a recorded human approval.
last actions
Queued behind flag atc_sticky at 50%review
Deployed #4189, first hour cleanshipped
Auto-reverted #4184 on error rate +0.8%closed
Rolled flag gift_guide to 100%shipped
Held deploy: no human approval on fileblocked
30-day
deploys
148
clean first hour
99.4%
auto-reverts
2
Human agentstage · gate
You
Every ticket stops here as a diff with the reasoning attached — what changed, why, what it is measured against, and what happens if it fails. Approve, edit or reject. You can auto-approve whole categories once you trust them, and revoke that in one click.
reads
The diff
The reasoning
The metric plan
produces
Approve
Edit
Reject
hard stop
Nothing ships past this gate without you, or a rule you wrote.
last actions
Approved sticky ATC at 50% exposureshipped
Edited copy on 4 of 46 descriptionsshipped
Rejected new drawer pattern, off-brandclosed
Auto-approve enabled for metadata changesshipped
Requested rollback rehearsal before Forge mergereview
30-day
median review
11 min
approved as built
94%
edited or rejected
6%
the handoff
Atlas sits here
→
→
→
→
→
→
→
→
→
Pick the four you want in the pilot.
Most stores start with Atlas, Muse, Volt and Sieve — scope, design, build, test — then add content
and data agents once the cadence holds.