the Mercury platform

Your AI finally
has a manager

Agents are drafting your emails, shipping your code, and rewriting your records, with nobody watching, nothing connected, and no memory of yesterday. AAOS is the missing management layer for the work AI now does. Every Mercury deployment runs on it.
See it in a deployment
Tour the platform
Web app + MCP servers
Model-agnostic
Runs on your cloud
WO-2401 / live protocol
Created
Planning
Ready
Dispatched
Image
Approval: human required
awaiting sign-off…
Image
In Progress
Review
QA
Done

The gate cannot be automated,
bypassed, or overridden. That's the point.

You adopted AI agents. 
Nobody is managing them

Three failures that compound quietly, every day, across every team that has let agents into real work.

No oversight

Agents act with no approval step. They push code, send deliverables, rewrite records. Enterprises build governance from scratch, or skip it and learn the cost later.
0 of 12 agents gated

No memory

Every conversation starts at zero. Client preferences, decisions, project history: gone the moment a tab closes. Intelligence evaporates from the organization daily.
0 KB retained context

No connective tissue

Each AI is a silo. Nothing synthesizes what's happening across projects, flags the risks, or tells you what to focus on now. The signal lives in the seams.
0 of 4 tools linked

A field report: one company trusted AI sales analysis for nine months.
The numbers were hallucinated. Budget, hires, and strategy, built on fiction. Nobody checked.
The AI was never wrong. It was never managed. This is why AAOS exists.

One workspace
where AI gets managed

A web app and MCP servers, with intelligence built into every screen. This is the system underneath every Mercury deployment. Not a slide. The actual product.
Flow
immediate · 09:00
Client escalation: EdgeCo invoice dispute
email + WO-2417 · response drafted
dispatch
before 2 pm
Approve deploy plan for reporting agent
WO-2419 · awaiting your gate
Vendor onboarding docs: 2 of 6 outstanding
agent chasing · no action needed
this week
Quarterly close readiness: on track, day 3 of 4
daily reconciliation clean
on track
01

flow / command center

Start every morning with clarity

Flow folds email, Slack, tickets, calendar, and active work orders into a single briefing ranked by urgency. Then it hands you action buttons to dispatch work from the page itself.

The briefing, not the inbox, is where the day begins.

Intelligence
Investor momentum building: 3 warm intros this week
signal · confidence 0.87 · email + work orders
evidence
Support SLA breach likely on ClientCo by Friday
risk · confidence 0.91 · tickets + WO history
act
Design work orders underestimated 35-50%
pattern · confidence 0.87 · 14 observations, 3 weeks
evidence
02

the intelligence layer

Intelligence that finds you

The intelligence layer surfaces signals, risks, and patterns, each with a confidence score and source attribution. It does not guess; it shows its evidence.

The synthesis that would take hours arrives already assembled.

Plan
Phase 01
Data pipeline hardening
3 work orders · complete
Phase 02
Backtest engine optimization
marked depends_on 01 · no shared inputs found
Phase 03
Reporting surface
depends_on 02 · confirmed
03

plans / dependency graph

AI that sees your critical path

Plans organize work into sequenced tracks with dependency chains. The system reads that structure in real time and notices when a task does not actually depend on the step before it. One click to parallelize, and your timeline shortens.


It catches what a spreadsheet quietly hides.

Intelligence
WO - 2401
approve
Auth middleware refactor
work order · in review · requires sign-off

acceptance criteria

Session tokens rotate on privilege change
All routes behind middleware pass integration tests
Migration plan approved before database change

Full audit trail
6 events · every approval carries recorded rationale
04

work orders / governed records

Every task is a governed record

Not a chat message. Not a prompt. A structured, traceable record: the objective, acceptance criteria, the assigned agent, dependencies, and the full trail of what happened.

The atomic unit of governed AI work.

Every piece of work moves through
the eight-state machine

The Dispatched to In Progress transition requires explicit human approval. It cannot be automated, bypassed, or overridden. That one property separates AAOS from every framework that just lets agents run.
Created
Planning
Ready
Dispatched

Approval:human

In Progress
Review
QA
Done

Created

Planning

Ready

Dispatched

Approval:
human

In Progress

Review

QA

Done

Gates are business decisions
Placed during design at every sensitive action: payments, external communication, data changes. Decided by you, documented, changeable any time. Opening a gate is a choice you make when the log has earned it.
Gates are granular
Per action type, per threshold, per workflow. "Approve invoices under $500 autonomously, route everything else" is a one-line rule, not a custom build.
the intelligence layer, underneath

Memory that thinks.
Not storage that recalls

The insight stream above is powered by a consolidation engine that synthesizes events into knowledge on a schedule. Every conclusion carries provenance down to the raw events.
observations / real-time

Raw events

Timestamped, source-attributed capture of everything that happens.
"WO-2401 completed in 6 hours. Original estimate was 4."
fragments / hourly synthesis

Patterns

Related observations clustered into confidence-scored patterns.
"Design work orders consistently underestimated by 35-50%. Confidence 0.87, based on 14 observations across 3 weeks."
narratives / weekly, permanent

Conclusions

Timestamped, source-attributed capture of everything that happens.
"Estimation bias is structural. Recommend a 1.5x multiplier on design work. Pattern has held 11 consecutive weeks."
Month three is better than month one, and nothing about your operations lives solely in one person's head. Including ours.

Every action.
Every decision. 
Every time

All agent activity is logged with full traceability: what was done, under which work order, past which gates, on whose approval. Immutable, timestamped, exportable.

Most clients discover something unexpected: agent-executed work is easier to evidence than the manual process it replaced. People forget to document. The system can't.

audit trail / WO-2401 / auth middleware refactor
10:15 AM
Created

Work order WO-2401 created — "Auth middleware refactor"

user: joshua

10:16 AM
Planning

Plan generated: 5 steps, 3 files modified

agent: senior-dev

10:18 AM
Ready

Plan locked. Acceptance criteria: 5 criteria defined

system

10:20 AM
Approved

"Plan looks good, proceed with refresh token implementation"

user: joshua — rationale recorded

10:32 AM
In Progress

Action: npm install jsonwebtoken

agent: senior-dev

10:45 AM
Gate — Pending HUMAN REQUIRED

prisma migrate deploy flagged as HIGH RISK

agent: senior-dev — escalated

Built to sit on top of 
what you already run

AAOS connects to your systems through APIs and the Model Context Protocol, the open standard for connecting AI systems to tools and data. Your ERP, CRM, document stores, and communication tools stay exactly where they are. Agents work through enumerated, permissioned connections listed in the design document and never expanded without sign-off.
Model-agnostic by design: AAOS orchestrates the models, it isn't one. Providers are swappable as the landscape moves, and deployments can run on your cloud under your provider agreements.
We run our company on it. Mercury's content operations run as AAOS work orders: drafted by agents, gated by humans, published with a full trail. The page you're reading moved through the same states.

Questions,
answered.

Today, AAOS ships as the platform under every Mercury deployment. If you're interested in running it independently, talk to us; the architecture was built for that future.