Brief ingested once - CF-B-0114. I'll keep you posted in plain language.

Account access
Create your workspaceLoginThe workspace begins after your kickoff brief is handed over.
Assistant Chat
You speak only with your Project Personal Assistant, and it is the only agent that speaks to you. Frontend, Backend, Integration, QA, Docs and the Orchestrator never contact you directly — they route questions and hand-offs between themselves. Anything they resolve on their own is written down as a logged assumption in the pinned ledger beside this conversation, so the reasoning stays visible even when nobody asked you.
Write to your Personal Assistant in plain language. It is the only place the conversation starts.
Work items moving between agent roles are reported back here in plain language.
Every assumption and decision logged by the agents sits in the ledger beside this conversation.
Only business-critical decisions are raised. Each arrives with a recommendation attached.
Brief ingested once - CF-B-0114. I'll keep you posted in plain language.
Can you confirm the pricing logic for annual plans?
Progress update - CF-0142 (Backend) moved to QA; CF-0143 (Frontend) in progress.
One decision needs you - pricing logic for annual plans. Recommendation attached.
One decision needs you - pricing logic for annual plans. Recommendation attached.
Bill annual plans monthly at the standard monthly rate, reviewed at the next natural checkpoint.

Account access
Create your workspaceLoginThe workspace begins after your kickoff brief is handed over.
Agentic delivery platform01 / ClientForge
Seven agent roles plan, build, test and document through one shared project board. Every assumption is logged, every decision stays traceable, and only the Personal Assistant ever asks you anything.
Four stops. One question asked only when it changes the outcome.
Structured project brief fed once after the human kickoff call.
Seven agent roles build, test, and refine continuously through the shared project board.
The Personal Assistant surfaces only high-impact decisions, with a recommendation.
Delivered faster than traditional human-only teams with full traceability.
Seven agent roles do the work. They communicate through the shared project board and structured messages, request clarification from each other or from the Orchestrator, and log every decision and assumption. No agent goes directly to the client.
Owns plan, timeline, dependencies, status; creates and prioritises tickets; decides which agent handles a task; monitors progress and unblocks agents; decides when something is truly blocked.
Builds frontend work items assigned by the Orchestrator.
Builds backend work items assigned by the Orchestrator.
Handles full-stack and integration work when needed.
Tests work items; returns failed tickets to the owning agent.
Produces documentation and handover material.
The only agent the client talks to; the single gate to the outside.
The project brief is fed once. After that, every agent reads from the same long-term store before it asks anything — so the client is never asked twice. Vector recall finds the passage; the structured tables hold the binding version of it.
Brief passages embedded once and recalled by meaning, so an agent finds the constraint that already covers its question instead of raising a new one.
The binding record: the brief items, constraints and logged assumptions a decision must respect. What the agents decided is written here, not re-litigated later.
| Table | Cols | Rows | Write |
|---|---|---|---|
| brief_items | 5 | 5 | canonical |
| constraints | 4 | 4 | canonical |
| tech_choices | 6 | 6 | canonical |
| success_metrics | 4 | 3 | canonical |
| assumptions | 7 | 7 | append-only |
| decisions | 7 | 7 | append-only |
| escalation_gates | 5 | 3 | rare |
RecalledRequirements, constraints and tech preferences are read from the store before any agent asks anything.
DecidedWhere the brief is silent, the Orchestrator decides and logs the assumption with the ticket that triggered it.
Never re-askedA question only reaches the Personal Assistant if it is business-critical and cannot be inferred or decided.
Agents create, assign, update and close tickets on the shared board without routing through a human. Each ticket carries its ID, its role line, its owner, its state and the token cost it took to get there.
The Orchestrator turns a plan step into a ticket with an ID, an owner role and a cost budget.
Orchestrator 01The ticket moves onto a role line — frontend, backend or integration — with the brief attached.
Role line 02Work lands against the ticket in the shared board; assumptions are logged as they are made.
QA 05Tests pass, the change is recorded in the audit trail, and the next ticket is pulled from the queue.
Docs 06Token cost is recorded per feature and set in tabular figures so the column reads like a timetable. Routine tickets are routed to smaller models; only architecture and complex reasoning escalate to the larger ones.
A work item reaches a human only when it leaves these three branches. Everything else is answered from project memory, decided by the agents with a documented assumption, or deferred to the next checkpoint.
A critical product decision cannot be reasonably decided by the agents from the brief or project memory.
ExamplePricing tiers and discount rules for the ordering flow.
Technical risk is high — authentication, data integrity, or a major architecture choice is at stake.
ExampleWhere customer records are stored and who can read them.
The client has given conflicting or incomplete information that cannot be resolved from project memory.
ExampleInvoice exports required as PDF and as "no file exports at all".
The brief asks for self-serve account export, but storing customer records outside the managed database conflicts with the data residency constraint in the brief. The agents will not resolve this one by assumption.
Keep records in the managed database in the EU region and deliver export as a queued job that is emailed to the account owner.
OrChoose differently and the Personal Assistant logs the decision, notifies the Orchestrator, and reassigns the affected tickets.
Version one — everything a service company needs on day one
The smallest build that is still a working system: five parts, shipped in order.
Each part below is already scoped to the kickoff brief. Nothing waits on a reply thread, and nothing is decided silently — every assumption is written to the project record.
Seven numbered agent roles with one Orchestrator owning plan, timeline and priorities. Agents talk through the shared project board, never to the client, so nothing routes around the Orchestrator.
A vector index for reasoning and structured tables for project facts, so a known answer is never asked twice. Context compression keeps routine work on smaller, faster models.
Agents create, assign, update and close work items on a ruled board. Every ticket carries an ID, an owner, a state and a token cost, and QA returns failed tickets to the owning agent.
One persistent assistant is the only agent the client talks to. It asks only when an answer is business-critical, always offers a recommendation, and reports progress in plain language.
Humans enter for critical business logic, high technical risk, or a brief that contradicts itself. Everything else is answered from memory or decided by agents with a logged assumption.
Starts after one call
In the studio
One project lead, one screen, one board. The agent workforce handles the back-and-forth, so the person who scoped the work is free to review it instead of chasing it.
A delivery lead reviews the day’s tickets before the client has opened the app. Work started overnight, assumptions are logged, and the only thing waiting is a decision that actually matters.
No comments yet. Be the first!