How It WorksField Brief  Nº 002

Your business, briefed one decision at a time.

LiquidSilicon opens on a briefing, not a dashboard. It greets you, counts what needs you, and deals your live business one card at a time — each carrying why it surfaced, when it’s due, and how it’s asking you to answer. Reply in a tap or just talk back. When a card is a problem, an agent drafts the fix and shows you the diff — and applies nothing until you confirm. Every decision files itself to a status board and a history you can audit.

N01One at a time

a briefing, dealt card by card — not a wall of alerts

N02Why · When · How

the grammar every item carries, so you decide from the card

N03Tap or talk

one-tap decisions, or barge-in voice that listens

N04You confirm

agents draft fixes; nothing applies without your yes

d1d2d3d4d5ROOTLEAFTRAVERSAL PATHMAX DEPTH 5
Fig. 1The knowledge field your console reads from — your operation as a live lattice of nodes and relationships. Each morning briefing is a traversal of this, resolved into the few decisions that actually need you.
01 / THE BRIEFING LOOP

Good morning, then one card at a time.

Step 01 · Greet → deal → answer → file

The console opens with a spoken “Good morning — three items need your attention,” then starts dealing. One card holds the center; you handle it; the next deals in. The queue tells you how many remain, never all at once. Everything that doesn’t need you is triaged in the background — and if nothing is urgent, the briefing says “all clear” and steps aside.

The point is your attention, not your engagement. A briefing is a short list of decisions, not a feed to scroll. When it’s done, you’re done — there is nothing left to go check.

Each card is narrated as it lands and settles with a spring — the shipped console motion, not a mock. You can turn narration off entirely and read; the flow is identical.

02 / THE EXPRESSION GRAMMAR

Every card tells you why, when, and how to answer.

A briefing card is not a headline you have to go investigate. It carries its own reasoning in three terse lines, so a decision can be made from the card itself.

WHYOne line of provenance — the reason it surfaced now. No alert ever arrives without telling you what tripped it.
“5 denials is 4× your daily baseline.”

WHENThe temporal context from real timestamps — recency plus any hard deadline, pulled out as its own element rather than buried in prose.
“Detected 07:42 · ongoing.” “Due Aug 1.”

STYLEHow the card is asking you to respond, stated up front — so you know the kind of answer it wants before you read the detail.
DECIDE · APPROVE · REVIEW

03 / ANSWER IT YOUR WAY

Talk over it. Tap once. Either works.

You get a briefing card and two honest ways to respond — your voice, or a single tap that moves the real record.

01

Barge-in voice

The briefing narrates itself, and the moment you start speaking it stops and listens. A live transcript fills in as you talk; your reply routes straight into the conversation. No wake word, no waiting for it to finish — you interrupt it the way you would a person. Where a browser can’t do speech, the mic simply isn’t offered.

02

One tap moves real work

The buttons on a card are honest. Approve approves; Advance moves a lead a pipeline stage; Send sends — each on the actual record, and it tells you what it did (“Approved.” “Sent.”). A button that only opens a record to read says so, and leaves the item in your queue unhandled. No button pretends.

04 / AGENTS DRAFT THE FIX

It doesn’t just flag the problem. It drafts the answer.

When a card is a problem — a spike in denials, a broken access policy — it can hold its own solution. The Builder drafts the actual change and renders it as a readable before → after diff with a risk note. You confirm & apply, or reject. Nothing is applied without your explicit confirm, and every apply writes a receipt.

AGENTPROPOSALOUTPUTSHIPSHUMANGATEPROPOSEAPPROVE ✓REJECT · SENT BACKNOTHING SHIPS WITHOUT A HUMAN YES
Fig. 2The circuit: a drafted fix meets the human gate — confirm & apply writes the change and a receipt; reject loops it back to the agent for revision. Guardian screens it either way.

Propose → confirm → receipt. The receipt is an attributed, timestamped record of exactly what was applied — part of the same append-only history everything else files into.

Never silently auto-applied — and we say so plainly, because the honesty is the product. An assistant that drafts fast but applies only on your yes is one you can actually leave running.

05 / EVERYTHING FILES ITSELF

Every decision has a place to go.

The desk reads like a real workspace, not one undifferentiated “handled” pile. Records you act on move across a live status board; the decision itself files into a history you can fan open and audit.

APPEND-ONLY EVENT LOG← LATESTE1E2E3E4E5E6E7REBUILDPROJECTION · ORDERSPROJECTION · REVENUEREPLAY FROM ANY POINT
Fig. 3The append-only log rebuilds its projections and can replay any past state — the same log is the filed decision history and the audit trail.

A spatial status board. Record-bearing items move New → In Progress → Done off to the side. Advance a deal and its card flies to In Progress; approve a payment and it lands in Done. You can see the state of the operation at a glance.

A macOS-style history. Every decision animates into a fan stack in the corner — filed by when you decided and what you chose. The console is its own record, so an answer is never lost.

06 / THE FIVE SPECIALISTS

Five agents assemble every briefing.

There is no single, opaque “AI.” Five narrow agents each do one honest job — one watches, one ranks, one drafts, one guards, one learns — and the two that could change anything, Builder and Guardian, are the two you have the most control over.

OBS-01OBSERVERWatches everyevent stream.ADV-02ADVISORRanks & coststhe options.BLD-03BUILDERWrites theapproved change.GRD-04GUARDIANVetoes anythingunsafe.LRN-05LEARNERFeeds outcomesback in.REGISTER OF AGENTS · N05ONE APPROVAL GATE
Fig. 4The five agents as a specimen sheet — eye, compass, pen-nib, shield, and cycle, each drawn in the same hand.
OBS-01ObserverWatches your event streams and names what changed — a stalled deal, a broken pattern, a spike. This is the signal that becomes an item in your briefing.Reads events. Never writes.Observe
ADV-02AdvisorTurns that signal into the expression you actually read — ranked, costed, with the why, the when, and a confidence score attached.Recommends. Applies nothing.Suggest
BLD-03BuilderDrafts the actual fix behind a card — schema, logic, interface — but only ever as a proposal you confirm. It cannot ship its own work.Writes proposals only.Automate
GRD-04GuardianScans every drafted fix for PII exposure, prompt injection, and policy breach, and can block it before you confirm.Scans and can veto.When enabled
LRN-05LearnerFeeds each decision you make back into the loop, so tomorrow’s briefing is ordered against what actually happened, not what was predicted.Closes the loop.Observe
07 / AUTONOMY YOU DIAL UP

You set the temperature.

Start with the agents drafting and you confirming everything. Autonomy is a dial, not a switch — five tiers, set per class of change. Agents earn the next tier by getting the last one right, and you decide when to grant it.

OBSERVE00SUGGEST01AUTOMATE02EXPAND03REPLACE04CURRENTCOOL · OBSERVE-ONLYHOT · FULL AUTONOMY
Fig. 5Autonomy as a laboratory temperature gauge, graduated observe → suggest → automate → expand → replace, needle on the current tier.
01
Observe

Agents watch and report. Nothing acts.

02
Suggest

Ranked, costed recommendations. Still no changes made.

03
Automate

Change classes you approved run on their own — Guardian still watching.

04
Expand

Autonomy widens to adjacent decisions in the same class.

05
Replace

The agent owns that workflow end to end, every action still logged.

Grant autonomy one phase at a time, per change class — and pull it all the way back to observe-only in a single click, with no redeployment and no downtime. Guardian keeps its veto at every tier, so raising the dial never removes the safety screen.

08 / HOW IT’S SET UP

It starts with a description, not a schema.

Your words become the runtime the briefing reads from

Before the first briefing, you tell LiquidSilicon what your operation actually is — the entities, the relationships, the rules — in plain language. From that it generates your config-driven runtime: the data model, the workflows, the screens, and the knowledge graph the agents reason over. No schema to hand-author, no rebuild to sit through. Change the description and the runtime reshapes to match.

TENANT APPYOUR OPERATION, DESCRIBEDSTANDARD PORTSMCPAPIEVENTSAGENT CLIENTCONNECTEDAGENT CLIENTvia APIAGENT CLIENTvia EVENTSONE BOARD · STANDARD PLUGS
Fig. 6Your operation, described, becomes a tenant app exposing standard ports — MCP, API, events — that the agents plug into. One board, standard plugs.
09 / WHAT IT RESTS ON

Three primitives, no magic.

The briefing is the surface. Underneath sit three things — an event-sourced core, a knowledge graph, and a config-driven runtime — each boring on purpose: auditable, replayable, inspectable. It’s also why the console evolves with your business instead of being torn out and replaced every few years.

01

Event-sourced core

Every decision you make in the briefing is an append, never an overwrite. Any past state is replayable and every action carries its own audit trail — which is why the filed history is the real record, not a summary.

02

Business knowledge graph

Your operation described as entities and relationships the agents reason over — not opaque tables they blindly mutate. The graph is the shared truth every briefing is drawn from.

03

Config-driven runtime

Each tenant’s software is generated from their own description of the business. Change the description and the runtime reshapes — no rebuild, no migration project. The console fits the operation instead of forcing the operation to fit the product.

10 / TWO SURFACES

One loop, seen from both sides.

You get a briefing; IT gets a control plane. The same decision travels one path — an item you answer in the Business Console becomes a proposal an agent makes, which an IT reviewer approves in the Control Plane before it ships. Two audiences, two screens, one loop.

ASKPROPOSEAPPROVESHIPBUSINESS CONSOLEDECIDEMORNING BRIEFINGCONTROL PLANEAPPROVEREVIEW · APPROVE · WATCHAGENTS ×5ONE LOOP · SEEN FROM BOTH SIDES
Fig. 7The loop in miniature: Business Console and Control Plane with the five agents between them, and the four arrows — answer, propose, approve, ship — that carry every change around the circuit.

These are faithful recreations of the console.

1You answerA one-tap or spoken decision in the Console.
2Agent proposesA scored, costed change — never applied yet.
3IT approvesA human accepts or rejects in the Control Plane.
4Change shipsIt deploys, logged as a replayable event.
End-user · Business ConsoleFig. 3
liquidsilicon.ai/console
LiquidSiliconConsoleITEM 1 / 3Briefing

Your agents

AI Assistant3 open itemsYour records and workflows
Email AgentStanding byDrafts, follow-ups, campaigns
Support Agent0 open casesCases, tickets, resolutions
Analytics AgentStanding byReports, performance, trends

Morning briefing · 3 items

Good morning, Sai.

Analytics AgentAnomalyDecide
Pipeline’s up — but one signal doesn’t fit.

Applications rose 22% while ad spend held flat — the two normally track. Worth a decision before you read the growth as real.

OPEN PIPELINE

$4.2M

+12%

WIN RATE

31%

+3%

STALLED > 30D

9

+4

FORECAST

$1.3M

+8%

WHY · applications normally track marketing spend — this break is unusualWHEN · detected 06:15 · 4-week window
InvestigateDismiss
New2
In progress1
Done4
Fig. 3The end user never opens a report builder — the dark console deals them one decision at a time, each with why it surfaced, when it’s due, live figures, and one-tap (or spoken) ways to answer. Answered items file themselves into the board below. Illustrative recreation — figures are examples.
The proposal crosses to IT — nothing has changed yet
IT · Control PlaneFig. 4
liquidsilicon.ai/control/approvals
APPROVALS QUEUE1 pending

Review AI-generated suggestions

🐢 AI-Detectedpatternhigh impact86%

Add index to slow orders query

Builder proposes a migration to index orders(created_at). p95 on the orders endpoint exceeded 1.8s across 2,412 events.

agent:builder3m ago
RejectApprove
Trust tierObserveSuggestAutomate ◂ currentExpandReplace
GUARDIANBlocked8m ago

Halted a Builder change: PII (email address) detected in a generated export template. Sent back for revision before any deploy.

Fig. 4IT sees every proposal with its confidence, evidence, and blast radius — approves or rejects it, and, when Guardian scanning is enabled, sees it flag anything unsafe. Illustrative recreation — figures are examples.
11 / QUESTIONS

The things people actually ask.

Q01

What actually happens when I open the console?

It greets you and tells you how many items need you — spoken aloud if you want, read if you don’t — then deals them one at a time. Each card carries why it surfaced, when it’s due, and how it’s asking you to respond, so you can decide from the card itself.

You answer in a tap or by talking back, the card files itself to a history and status board, and the next one deals in. When nothing is urgent, it says “all clear” and gets out of your way.

Q02

Can the AI change things without approval?

By default, no. When a card holds a fix, an agent drafts it and shows you a readable before → after diff with a risk note — but nothing is applied until you confirm. When you do, you get a receipt: an attributed, timestamped record of exactly what happened.

You set an autonomy tier per class of change; only the classes you deliberately raise run on their own, and even then, when Guardian scanning is enabled, it screens each one and can veto it. You can pull the whole system back to observe-only in a single click.

Q03

Do I have to talk to it?

No. Voice is an option, not a requirement. The briefing narrates itself and listens if you speak — start talking and it pauses, fills in a live transcript, and routes your reply into the conversation. Every card is equally answerable with a single tap.

Voice degrades gracefully: where a browser can’t do speech recognition, the mic simply isn’t offered and nothing else changes.

Q04

Do I need my own AI API key?

LiquidSilicon is bring-your-own-key first. Connect a key from Anthropic, OpenAI, Google, DeepSeek, or Mistral and every agent call runs under your own provider account — your prompts and data are governed by the agreement you already hold with that provider, not by us.

If you would rather not wire up a key on day one, you can start on a capped starter allowance we provide and switch to your own key whenever you like. Either way, you always see exactly what each agent spends.

Open it tomorrow morning to a briefing.

Point LiquidSilicon at your operation and let the work find you — the few decisions that actually need you, each ready to answer in a tap or a sentence, with the agents drafting and you confirming.

Free tier · Humans confirm every change