From the team behind Amazon's foundational AI
30 yrs in enterprise

An AI agent that does the work,
not just the chat

Today's AI tools draft and advise, then hand the work back to you. Alient is the agent that takes the whole job — joins the meeting, operates your tools, remembers every decision, and delivers it done.

72.1%
#1 on OSWorld Benchmark
+9.2 pts
vs. previous SOTA (Sonnet 4.5)
Beats
Claude Code · Terminal Benchmark
Zero egress
On-prem · VPC · managed
alient@mac · session live · 09:41
$ alient plan JIRA-4812
→ reading repo… /src/checkout/totals.ts
→ planner: 3 steps · 12 files
→ coding… totals.ts +42 −18
→ tests ✓ 214 passed · 2 flaky
→ reflector: ok · adversarial verify ok
$ gh pr create
opened #1284 fix: prorate taxes on split ship
A
alient
now
Opened #1284 · fix: prorate taxes on split ship. CI green. Ready for review, @maria.

Watch Alient pick up a real task and finish it end to end — one example of the agent completing work. Recorded April 15, 2026.

01 The shift

You don't need another chatbot. You need an agent that does the work.

Every AI tool you've tried generates drafts, suggestions, summaries — and leaves the work on your team. Alient is the first one with its own machine, its own logins, and its own accountability. It picks the task up, finishes it, and tells the team it's done.

Chatbots & copilotsAlient
Where it runsA chat window in your browserA dedicated machine — on-prem, VPC, or managed
IdentityActs as youIts own Slack, email, logins
Data pathYour data → vendor cloudStays in your private deployment
OutputDrafts a human finishesFinished work, end to end — code, follow-ups, memos, updates
Why Alient

A different species of agent.

#1 on OSWorld computer-useOperates any app — no API needed5-type compounding memoryPrivate deployment · zero egressAccountable autonomy — audit & rollbackArchitectural moat — not a wrapper

Chatbots & copilots

You operate them. They draft, suggest, and summarize — then the work comes back to you to finish and ship. Alient operates your tools itself and returns finished work.

Cloud “AI coworker” bots

They live in a chat window, reach only what has an API, and run in a shared cloud. Alient reaches any app through computer-use, runs single-tenant on a machine dedicated to you, and is accountable for the result.

Frontier model labs

They sell horizontal capability. Alient is the deployed, private, accountable agent that runs on top of their models — and bundles computer-use, memory, safety, and deployment into one agent.

If it can autonomously ship production code — the hardest knowledge work there is — it can run the rest of your operation.

The same engine already runs Ontfly, our consumer product where the agent builds the website, runs the ad campaigns, drafts customer replies, and manages sales for solo founders.

02 How it works

A real workday, automated end to end.

01 · Memory

Remembers everything.

Five-type persistent memory — procedural, episodic, semantic, working, declarative. Your business, your conventions, every decision in every meeting. The work gets sharper every week.

see the trace
02 · Meetings

Joins and talks in meetings.

Dials into Zoom, listens through a virtual mic, speaks when relevant, asks clarifying questions, and captures every action item — then delivers them after the call.

see the trace
03 · Tools

Operates any tool.

Works the apps you already use, including the ones with no API — portals, CRMs, finance tools. It clicks, reads, and types like a person.

see the trace
04 · Delivery

Delivers finished work.

Picks up a task and returns it done. The proof: it opens the codebase, writes the change, runs the tests, and ships the pull request — the hardest version of finishing a job.

see the trace
03 Proof

We didn't ship a chatbot. We shipped the #1 agent on the world's hardest computer-use benchmark — it operates software, it doesn't just talk about it.

72.1%
#1 on OSWorld Benchmark · March 2026.
View leaderboard →
#1
Among all general models on the benchmark.
Methodology →
Beats
Claude Code on Terminal Benchmark, same Claude Sonnet 4.6 backbone — architectural advantage, not model advantage.
Read the white paper →
OSWorld Benchmark · Jan 2025 → Mar 2026Higher is better
80%60%40%20%Jan 25Jun 25Nov 25Mar 26Previous SOTA · 62.9%ALIENT · 72.1%
The win is architectural. A 5-type memory system, a Planner→Programmer→Reflector graph with adversarial verification, and a two-layer safety policy on every tool call. The same model in a different system gets a different answer.
Read the architecture →
Get started

See what Alient can do for your team.

04 Architecture

Seven layers. One agent.

The computer-use agent is the general executor — it operates any app on its machine, no API required. The coding and meeting agents are specialized executors beside it, sharing the same memory, safety policy, and audit trail.

01 · Event busTemporaldurable workflows02 · OrchestratorLangGraph · Planner → Programmer → Reflectoradversarial verify03a · GUI agentComputer useclick · type · read screenany desktop app03b · Coding agentRepo-nativeworktree · tests · PRCursor · terminal03c · Meeting agentZoom · Meetvirtual mic + speakerdiarized transcript03d · ReflectorAdversarialself-critiquerollback on red04 · MemoryFive types · persistent across sessionsproceduralepisodicsemanticworkingdeclarativeGraphiti · memU · A-MEM · files05 · SafetyTwo-layer policy on every tool call06 · ObservabilityLangfuse · full traces · replay
Authored SVG · seven-layer system map
05 Security

Private by design. CISO-ready by construction.

Runs where you choose.

Your own hardware, your VPC, or a dedicated managed instance we run for you. Always single-tenant — never a shared cloud. Under your MDM and DLP when it's inside your perimeter.

Zero data egress.

Reasoning, memory, screen capture, and audio all stay on the agent's machine. Only the LLM API call leaves — and even that can be self-hosted.

Graduated autonomy + audit trail.

Every action is classified by risk tier. High-risk actions require a human via #agent-escalation. Every tool call is logged to Langfuse with full traces.

SOC 2 Type 1 · targeted Q3 2026GDPROn-device LLM option · Apple SiliconISO 27001 · targeted 2027
06 Economics

The economics of an AI hire.

Priced against the role it replaces, not against tokens.

A full-time hireAlient
Fully-loaded cost$180,000 – $360,000 / yr$12,000 / yr
Ramp to first finished task30–90 daysDay one
Hours active40–50 / week24 × 7
SeveranceYesNone
Memory of decisionsUntil they leavePermanent

$12,000 per agent per year ($1,000/mo). Flat rate — no token metering, no setup fees. Design-partner terms available.

07 Letter from the founders

We built the agent, not the tool.

Between us we've shipped software at scale for two decades and grown up inside it for almost one. We watched the same pattern from both ends: every team has more work than people, and everyone has more meetings than focus time. The tools we tried — copilots, assistants, chatbots — moved the bottleneck around without removing it. The work still landed on a human.

So we built the agent instead of the tool. Alient is the colleague we wished we had: it picks up the work that used to land on a human — joins the meeting, asks smart questions, returns it finished. It is not infrastructure. It is a hire.

If that resonates, we'd like to show you a real session — not a slide deck. Thirty minutes, your stack, your task, our agent.

— The founders, Austin, TX · April 2026

Read our research →
CTOY.K.

Yurii Kulaksyz

Co-founder & CTO · 8 yrs experience

Built Alient's agent architecture and computer-use layer. Reads commit histories the way other people read novels.

CEOT.S.

Thirunavukkaresu Swaminathan (Swami)

Co-founder & CEO

Principal engineering manager. Two decades shipping enterprise software and security systems at scale.

Backed by[angel roster TBD]
09 FAQ

Objections, answered plainly.

Isn’t this just an AI chatbot or a Slack “AI coworker”?+
Chat tools and coworker-bots live in a chat window and reach only what has an API — you still drive them. Alient operates your actual tools, including the ones with no API, runs in a private single-tenant deployment, and is accountable for the finished result. And it's the #1 computer-use agent on the OSWorld Benchmark — not a wrapper around someone else's chat.
Why won't OpenAI or Anthropic build this?+
They sell horizontal capability. Alient is a deployed, private, accountable agent — not a model — that bundles computer use, memory, safety policy, and enterprise deployment into one SKU, and it runs on their models. The moat is architectural and operational, not weights.
What happens when it makes a mistake?+
High-risk actions are gated by a two-layer safety policy and escalate to a human in #agent-escalation before running. Low-risk mistakes are caught by the reflector and by CI; when they slip through, the same audit trail that tracks the PR tracks the rollback.
What kind of work can it own?+
Bounded, repeatable knowledge and operations work with a clear definition of done — a change shipped, a follow-up sent, a record updated, a report delivered. For ambiguous judgment — product decisions, cross-team negotiation, novel design — it escalates to a human. On coding work, the hardest case, a single Alient closes the equivalent of 0.8–1.2 junior-engineer tickets per week on our design partners' backlogs.
How does it handle work with strict security requirements?+
Deploy it inside your perimeter — on your own hardware under your MDM and DLP, or in your VPC. Nothing leaves the deployment except the LLM API call, which can be pointed at an on-device model, a VPC-private endpoint, or a self-hosted model you control.
What languages and frameworks does it support?+
Anything its machine can run. We've deployed into TypeScript, Python, Go, Rust, Swift, Kotlin, and Ruby stacks. The coding agent is framework-agnostic; the computer-use layer means it works with your existing IDE, terminal, and review tools — not a replacement for them.
Does it require any infrastructure on our side?+
Only if you want it to. Run it on a machine you provision — a Mac on your network, under your MDM — or in your own VPC, or have us run a dedicated managed instance for you. Either way we send a setup script and a checklist; day-one setup is under an hour.
Can we self-host the LLM?+
Yes. We support pointing Alient at an on-device model, a VPC-private Bedrock or Azure endpoint, or a self-hosted open-weights model. For the most common deployment today the endpoint is Anthropic with a zero-retention agreement.
How is pricing structured?+
$12,000 per agent per year ($1,000/mo). Flat rate — no token metering, no per-seat surcharge, no setup fee. Design-partner terms available through Q3 2026.

Put Alient on your hardest workflow.

30-minute intro call. We'll walk you through how Alient works.

We'll reply within one business day. No drip campaigns, no shared lists.