How do you deploy AI agents in production?

Deploying AI agents in production means more than hosting a model loop. The agent needs an isolated place to run, an authenticated identity, credentials scoped per user, memory, policy that can refuse an action, and a trace of what it did. Kodeus supplies that operating layer, so you describe the outcome and it handles deployment, whether self-hosted in your VPC or managed by us.

Does my data leave my infrastructure?

No. The runtime executes against your own database, under your own network policy. Runtime conversations, user data and credentials stay inside the environment you deploy into.

Am I locked into one model provider?

No. Models are selected per agent and swapped in a line of config. Cost-, latency- and policy-based routing with provider fallback is maturing on top of that.

Is this a framework, or the infrastructure under one?

Both, and that is the point. A framework helps you compose prompts and tools, then leaves you to build the runtime, identity, policy and observability yourself. On Kodeus you author the agent as a spec, the platform drafts it, and the same platform runs it: lifecycle, per-tenant concurrency, credentials, policy, traces and metering. It is where the agent gets built, not only where it gets hosted.

Can I use my own models and tools?

Yes. Models are chosen per agent and stay interchangeable, and any MCP server can be attached with one command. Your agent can also be exposed as an MCP server itself, or talk to other agents over agent-to-agent transports. If you already have a product, the CLI and SDK let you embed Kodeus operating services inside it.

What will it cost to run an always-on agent?

On the planned rates, roughly $58 a month in runtime before model tokens, and nothing at all if you self-host. Usage is metered at the execution level with model and tool consumption captured per agent, so you can see what each one costs before the invoice does. See pricing for the full breakdown.

Not yet. Kodeus is in private preview while we finish the SDK and the demo app. Join the early-access list and we will get you in touch with a build as soon as there is one, starting with teams whose use case we can support properly.

The Operating Layer for AI-Native Applications

Your product's agent needs an identity, its own credentials, memory, guardrails and a record of what it did, for every user. Kodeus is the runtime that gives it those, so your team ships the product, not the platform under it.

Request early access →

Your application3 users

{{ u.name }}

Kodeus runtime live

turn{{ reqLine }}

          {{ l.t }}{{ l.who }}{{ l.msg }}

        

Modelauto-routed

Any model, auto-routed per task.

    per-user identity·three credential stores·45 guardrails·five checkpoints·OpenTelemetry traces·spend caps·loop detection·approval records·open source on PyPI·
    per-user identity·three credential stores·45 guardrails·five checkpoints·OpenTelemetry traces·spend caps·loop detection·approval records·open source on PyPI·
  

Why agents stall before production

The model is not the problem. Everything around it is.

We know because it broke on us first. Four failures, and the moment each one shows up.

See how the runtime handles each one →

Kodeus{{ s.fix }}

Show, not tell

Watch the runtime refuse.

Every turn passes five checkpoints, including one before any tool call. If a rule cannot be evaluated the action does not happen and the turn goes to a person. This trace is what your compliance team sees.

        Console
        Approvals {{ pendingCount }}TracesAgentsPolicies

Held for review

{{ q.agent }} · {{ q.user }}{{ q.state }}

{{ q.sub }}

{{ detailState }}turn_7f3a91

support-agent wants to send_email on behalf of user_4821

Rule
claims.no_unverified_guarantee
Checkpoint
3 · before tool call
Reviewer group
support-leads
Spend so far
$0.0027

−returns are guaranteed for 90 days, no questions asked
+returns follow our standard returns policy

You are the reviewer.

Released by you · email sent with redaction · trace exported to OpenTelemetry

Turn trace
{{ r.t }}{{ r.tag }}{{ r.title }}

Open a live trace in the Console →

How it works

Define it. Run it. Operate it.

Read the quickstart →

CONSOLE

+ New agent

Agents

Tools and credentials

Knowledge

Skills

Schedules

Templates

support-agentrunningPlayground

support-agentResolves support requests in the systems of record. Refunds over the cap wait for a person.

ConnectorsZendeskGmailPostgres+

Skillsreturns-policytone-guide+

KnowledgeBilling policy 2026Product FAQ

Modelanthropic/claude-sonnet-4

Policyclaims.no_unverified_guaranteespend.cap $2/turnapprove: refunds.create

Localsame runtime

Your cloudselected · vpc us-east-1

Kodeus Cloudsame runtime

Airgappedsame runtime

Users2,310 isolated

user_4821own vault · own memory · own limitsready

user_0977own vault · own memory · own limitsready

user_1130own vault · own memory · own limitsready

deployed · 3 replicas · nothing calls kodeus.ai at serve time

Approvals1 pending

Behaviour tests17 / 18 pass · 1 diff

Spend$41.20 · 12,804 turns

Tracesotel://collector.acme

refunds.create · user_0977 · $640 over capHeld for m.chen. Tool did not run. Draft and verdicts attached.

diff · refund_over_500 · after model change to gpt-5Now asks for approval (was auto). Review before ship.

Proof, not a logo wall

Built on Kodeus, running with real money.

Warren is an autonomous fund manager we built on our own runtime and ran on live markets before asking anyone else to trust the platform. Every trade passed the same guardrails, approvals and traces your product would use.

The same core runs Nash for enterprise finance and embodied agent research with ARTPARK at IISc Bangalore.

How Warren is built →Case studies →

{{ n0 }}

traders

{{ n1 }}

paper volume

{{ n2 }}

live capital

{{ n3 }}

strategies · 16+ exchanges and prediction markets

Case studies

What already runs on it.

Three named runs. Warren is ours. Nash is enterprise finance. ARTPARK at IISc Bangalore is a research partnership. The industry cards on use cases are shapes, not customers.

All case studies →

Ours · live markets Warren An autonomous fund manager. Copilot waits for approval. Autopilot executes inside the limits the user set. A daily loss limit can hold an order. Enterprise finance Nash Long-running workers that monitor systems, reconcile records and escalate to a person when policy says they must. Research partnership ARTPARK Embodied agent research with ARTPARK at IISc Bangalore, on the same identity, policy and trace model.

What you get

Identity. Control. Evidence.

identity/

Identity and isolation

Per-user identity and tenancy. Three credential stores, encrypted per user with AES-256-GCM. One user’s connected accounts are unreachable from another’s session, and we test that with a six part adversarial isolation suite.

policy/

45

Control

Guardrails across six policy families at five checkpoints. Monitor, redact, block, escalate or hold. Spend limits and loop detection enforced, not prompted.

traces/

Evidence

OpenTelemetry traces with rule verdicts, latency and cost per call. Approval records with approver and time. Execution history you can hand to an auditor.

deploy/

Same runtime in all four. No dependency on Kodeus at serve time. Deployment guide →

Any model, auto-routed per task. OpenAI, DeepSeek and Claude today. The model should be replaceable. The operating layer should not be.

Developers

Read the code before you talk to us.

Open source SDK, one yaml file per agent, the runtime in Docker on your laptop. The Developers page has the terminal, the spec and the deploy path.

Quickstart Console

Questions developers actually ask

From the blog · 28 Sep 2026

How to deploy AI agents in production

Deployment is the same agent you already ran locally, placed on a boundary you chose. The policy engine does not get weaker on the way to production.

Read the post → All posts

Stop building the platform. Start shipping the product.

Tell us what you are building and we will get you into the preview as soon as we can support your use case properly.

Request early access → Book a call