Inside Briu · August 2026
The business runs through one conversation.
I built Briu's internal workflows so a conversation could carry work between tools without making me repeat every handoff. This August 2026 account shows the useful pieces, the controls and the unfinished parts. It is an internal proving ground, not a platform every customer needs.
What happens after I speak
The real problem
The AI was working. The owner was still the workflow.
The problem was not a shortage of capable tools. It was the repeated work of moving context between them: finding the right account, explaining the request again, gathering the result and deciding whether it was actually finished. A long-running agent session helped, but without durable workflows I was still carrying the handoffs.
AI can be everywhere while the workflow stays fragmented.
Research, spreadsheets, marketplaces, scheduled tasks, and browser work can each use AI and still remain separate systems. Every part gets smarter while the owner keeps carrying the handoffs between them, and the answer to that is not a large replacement system.
More agents can create more work at the center.
Agents can research and draft while their operator still moves the output, reviews the growing queue, and decides what happens next. More agents multiply that coordination. The better target is one bounded workflow arriving fully carried, with a person's attention required only where it changes the meaning.
That is the operating model: the person supplies direction, judgment, taste, and responsibility. The system supplies memory, retrieval, mechanics, follow-through, and evidence. Briu does not try to remove the owner. It removes the owner's job as the integration layer.
What got built
Three operating workflows and one design rule.
The first three describe Briu's business operations at publication. The fourth draws a design lesson from separate personal creative projects. A reusable method does not mean sharing business records or permissions with creative work.
From scattered history to a useful conversation.
Before a customer meeting I ask for the prep from my phone. The agent resolves the right account, reviews the available correspondence, checks current notes, researches the business, and hands me a private brief. During the meeting, the notes I type can flow to the same session for live research and record updates. After an on-site visit I sit in the car for ten minutes and dictate direction. The system works from my Mac at home while I drive, and by the time I walk in that evening the record is current, the next steps are concrete, and the outreach emails are drafted for my review.
From mailbox work to one governed review.
The same conversation reads across Briu's governed mailboxes, prepares a reply in the right voice, shows the exact sender, recipients, subject, body, and attachments, then stops. If I approve that artifact, a single-use authorization sends it and the provider receipt is read back.
From a spoken outcome to verified live work.
A software change becomes a bounded job with its files, constraints, acceptance checks, and required proof. It runs in a separate workspace, passes tests and review, reaches the live product only from the approved revision, and is checked after deployment. A session can keep working while I sleep. It cannot grant itself permission or declare itself shipped. A committed change is not itself proof of deployment, use or a business outcome.
Build only the missing layer.
Separate personal creative projects tested the same build decision. For a ski film, a phone-first editor supplied an interface that did not exist. For music, a voice layer drove Ableton rather than replacing the specialist tool. Every project, business or creative, presents the same choice: replace, connect, or wrap only what keeps the work from finishing.
The compounding effect
Every finished workflow lowers the cost of the next.
Useful pieces persist. A new workflow begins with an operating memory, connected sources, tested controls, reusable tools, and examples of what counted as done last time.
Direction persists
The request, decisions, and reasons survive the chat. The next agent starts briefed.
Sources stay mapped
A connected mailbox, CRM record, calendar, or repository stays available to every later workflow.
Controls get reused
Identity, approval, isolation, and receipt patterns are solved once, then reapplied wherever the same risk appears.
Repetition becomes a tool
Repeated mechanics become reliable code once the pattern is understood.
Proof improves the next brief
Results and failures become evidence. The next workflow inherits a sharper acceptance test.
Technical depth: the six operating contracts
The operating contracts
The same six questions govern every workflow.
These are design requirements across different workflows, not a claim that every path was fully implemented or verified.
One current front door.
Every session begins from one maintained picture of what matters now, where the facts live, and what remains unfinished.
Unknown stays unknown.
Email, CRM, calendar, code, and media each have an owner, read path, and permission boundary. A missing fact returns as a gap, not an invention.
Use intelligence before automation.
A frontier agent handles ambiguity and exceptions. Only well-understood repetition becomes deterministic code.
Package outcomes, not chat history.
Work travels with its outcome, scope, constraints, acceptance checks, and required proof, not a transcript.
Bind approval to the exact act.
A consequential action stops at a preview for a named person. The approval covers only that artifact and expires.
An answer is not a result.
Built, delivered, verified live, used, and outcome are separate claims. Work stays open at the first missing level.
The remote working setup at publication
How little friction it takes
The interface is a conversation, not a console.
Nearly every request I give the system is dictated. On the Mac I speak into a HERDR session with separate panes for the administrator and worker lanes. From an iPhone or iPad, a private remote connection returns me to that working session. The administrator assigns bounded work, keeps the source and proof visible, and coordinates the lanes while each worker completes its own step. The operating experience is simpler: I open the conversation, say what changed, and continue from the same place.
I have not needed to personally open Gmail to run the verified three-mailbox workflow in months. Reading, drafting, exact review, approved sending, and receipt checking happen without making me operate the mailbox.
Provider independence and the historical usage ledger
Why the work survives
The business owns what the agents learn.
Codex, Claude Code, and Cursor can all enter the same Briu environment. None is the permanent home of the company. Briu keeps the context, source map, tools, permissions, jobs, and proof, then chooses an available agent for the work. Changing providers does not require rebuilding the operating memory. A forthcoming note describes the coordinator, HERDR lanes, and bulletin that now organize this work.
How another business adopts it
Close one real workflow first.
Briu OS is the environment I built for Briu. Another business can borrow the method without adopting this internal system. Work with another business starts at the workflow or idea level, read-only, inside the operating day that already exists. The right first move might be coaching around current tools, a control layer that drives them, or a focused replacement for one weak link. If an existing tool is producing a reasonable return, changing it is not the goal; the workflow decides.
A bounded workflow is fully cooked when it begins from a clear direction, reads the right facts, carries the repeated mechanics, stops at the right person, and leaves a verifiable result. It is not fully cooked because six agents produced six drafts.
Observe the real day.
Read the existing sources and handoffs. Name what is proven, stale, or held in one person's memory.
Choose one expensive point of friction.
Pick the workflow that consumes attention, creates re-entry, or waits on one person's exception.
Improve it beside the people who know it.
Preserve what works. Use AI judgment for the messy part and tools only for clear repetition.
Keep authority with a named person.
Routine reversible work continues. Consequential actions stop at an exact review.
Measure use, then choose the relationship.
Continue through direct coaching, a bounded build, or both only where the evidence supports it.
Evidence boundary · August 17, 2026
What is real, partial, and not claimed.
The following is the dated publication snapshot, not a live status page. Used, partial and intended are different states.
In use at publication
- One maintained starting brief and durable company context shared across coding agents.
- Customer, CRM, three-mailbox communication, calendar, repository, and project-specific read paths with separate account boundaries.
- Governed email drafting, exact human approval, provider send, and receipt checking.
- Bounded job handoffs, separate engineering workspaces, tests, delivery checks, and result receipts.
- Dictated work from a Mac, phone, or tablet, with long-running sessions able to continue while I step away.
Real but still partial
- The first-in-command operating role works as a method; seamless resident continuity across every surface is unfinished.
- Voice dictation works as text input. The separate phone capture endpoint still lacks audio-only transcription.
- Autonomous work now has a zero-model release runner that readies merges, updates branches, verifies the merge chain, and wakes a worker only for a conflict or CI failure. Executor recovery and lifecycle cleanup still have named limits.
- The lifetime value ledger is not publication-ready because Claude, Codex, Cursor, direct, subagent, and dispatch coverage cannot yet be reconciled without named gaps.
- Software delivery is easier to prove than adoption. Real use and business outcome still require dated operating evidence.
Not claimed
- No unrestricted access to every account, application, customer, or business.
- No autonomous consequential sends, commitments, purchases, signatures, or sensitive access grants.
- No unlimited frontier-model usage or immunity from changing provider rates and limits.
- No promise that every existing workflow should become an agent or that current software should be replaced.
- No separately licensed Briu OS product, and no claim that a company can run without human direction, taste, or responsibility.
Current measurement practice
We read the active Codex accounts and Claude plan windows before assigning work, then record the source date, window, and limits. That live budget read is the current practice. The older retained-transcript count remains a historical account of one August week, not a measure of business value or cash spend.