Multi-agent framework
Built on a multi-agent framework: four AI agents, Clay, Greg, Jemma and Cody, share one conference line and one canvas. You talk; they answer by voice and build the work in front of you. Jev, the gatekeeper, scores and ranks their competing versions so the final call is yours.
Automation and AI enablement · solution architecture
Studio architecture · design & measured gates
How spoken intent becomes verified Salesforce work: combining the live conference canvas and measured release checks (as of 6 Nov 2026) with the target voice-binding and dual-plane bus design.
01 · Voice-to-agent binding (target design)
In the studio binding design, each persona on the line binds to its own working agent seat. A single-speaker floor lock keeps voices from talking over the room, while any irreversible action pauses for an explicit spoken read-back before it runs.
02 · Dual-plane event bus & state ledger (target design)
The target bus design separates speed from auditability: a fast event plane is designed to wake the assigned agent when a turn arrives (replacing polling loops), while an append-only state ledger records every handoff, enforces a design constraint of automatic token scrubbing, and feeds Jev to rank competing versions.
03 · Shared studio canvas & presenter seat (design)
While you speak, Clay, Greg, Jemma and Cody build OmniScripts, FlexCards, Data Mappers and trade-off cards in place on one canvas. For screen-share walkthroughs, the presenter seat (/present/) is designed around a role constraint: screen frames only, with no microphone, audio, or data writes.
04 · Takt release trains & measured break-it gates
Changes ship on scheduled release trains. Before any train departs, an independent agent reviews the exact version, automated break-it checks (1,457 chair tests and 1,215 break-it checks measured 6 Nov 2026) run against it, and the release guard confirms no live call is interrupted.
AI agent model
Each AI agent holds one accountable role. Four specialists build and challenge options in parallel, Jev ranks competing versions on a fast decision model, and a person makes every final call.
Session lead · implementation & release
Leads the working session, turns agreed decisions into working Salesforce and web changes, and drives the scheduled release train.
Strategy & decision pace
Frames the trade-offs, challenges scope creep, and keeps every working session moving toward a concrete decision rather than a growing backlog.
Architecture · adversarial reasoning
Stress-tests system boundaries before code is written and shapes how every diagram, card, and workflow reads for the people using it.
Delivery management · test & review
Writes the automated verification suites and inspects every proposed change at its exact version so no agent ever grades its own homework.
TypeSafe System One · ranking engine
A fast, low-cost decision model that scores and ranks competing versions on a team card so a heavyweight builder model is never the judge.
Accountable approval gate
Every ranked option, architecture trade-off, and release train waits for a person to make the call. No change ships on an agent’s word alone.
Multi-agent delivery model
How one idea becomes shipped work: decomposed, dependency-mapped and delivered by an agent fleet working in parallel on scheduled release trains.
01 · Envision
An engagement begins as a concept — a customer website, a portal, an automation. The intent behind it is recorded as a working charter: the outcome sought, the constraints that bound it, and the measures that will say it worked.
What works today
01 · Live conference line
You speak, and Clay, Greg, Jemma and Cody answer by voice. Their work lands on a shared canvas while you are still talking.
02 · Living cards
Say “make it shorter” or “make the logo bolder” and the card already on the canvas changes in place. No new copy to chase.
03 · Team cards
“Everyone, design a logo” gives one card with each agent’s version next to the others. You pick the one to keep.
04 · Discussion cards
Ask for a discussion and get the pros and cons with a summary. The agents see each other’s cards and build on them.
Salesforce focus
The agents build OmniStudio pieces, OmniScripts, FlexCards and Data Mappers, in a connected Developer Edition org.
Built, reviewed and released by the agents
Done means the person who asked has seen it work. Counts from the build’s own history, as of .
Chair tests
1,457 about 35× since the first day, 26 Sep
Break-it checks, every change
1,215 about 31× since the first day, 26 Sep
Each breaks one rule on purpose. One uncaught break fails the change.
Changes merged
90 this week, 2 to 6 Nov · 10 the week before
801 since 1 Sep, across the conference line, the board and this site.
Release gate
No release starts until every change has:
One release at a time. Deploys check that no call is live and keep a rollback snapshot.
The Integration Gap
Merged is not done
6 Nov: a fix was built, reviewed and deployed in about two hours. Its first live test found one more gap, within minutes, and that is being fixed before it counts as done.
Division of Labor