An operating system for agent fleets, built around an Agent Kernel.
A kernel that schedules, isolates, persists, protects and accounts for every agent, a language whose effects are the kernel's system calls, and industry packs.
Where the agent decides, and where the organization decides.
01
The agent chooses the path: which tool to call, which data to fetch, how to respond.
02
That part is probabilistic, and it's what makes the agent useful.
03
What it can reach isn't its choice: systems, data, models and spending limits are the organization's rules.
04
The rule is evaluated outside the agent, before the call goes out.
05
The agent decides within the limit. The limit still belongs to the company.
Framework, managed platform or operating system?
Who guarantees nothing is lost?
In a framework, each team's code. On a managed platform, the service, within its limits. In the AgentOS, the kernel, with a checkpoint at every step, and replay.
Where does the business rule live?
In a framework, in the prompt or the code. On a managed platform, in the vendor's policies. In the AgentOS, in a rules engine, versioned and explained.
How does it scale to thousands?
In a framework, through a separate infrastructure project. On a managed platform, by whatever the service charges. In the AgentOS, through actors and sharding, native to the runtime.
Who audits the agent, and who stops it?
In a framework, third-party tools. On a managed platform, separate consoles. In the AgentOS, the same kernel: passport, trail and kill switch.
And voice?
In a framework, it is another project. On a managed platform, another product. In the AgentOS, one more channel for the same agent.
How does the agent improve?
In a framework, with a new sprint. On a managed platform, with a separate optimization tool. In the AgentOS, the system itself measures, proposes and proves.
Where does the data live?
In a framework, wherever the team set it up. On a managed platform, in the vendor's cloud. In the AgentOS, it can stay in the customer's cloud.
Control, execution and build, on the same kernel.
Control Plane
Governance, policy, security, observability, auditability and FinOps.
Runtime
Execution, orchestration, parallelism and resource management.
Builder
Agent creation, workflows, tools and models.
Ecosystem
Enterprise systems and agents from other frameworks, which point their model traffic, tools and invocations at Tessera.
An Agent Kernel under every agent.
Agents
Renegotiation, claims, onboarding, customer service, and third-party agents under the same governance.
Language
Matiq, with a compiler, types and evaluation; Studio, CLI and debugger.
System calls
Every language effect, from a model call to a human approval, is a call to the kernel.
Agent Kernel
Scheduler, memory management, persistence and security.
System services
Model gateway, rules engine, event rail, audit, cost and value, voice.
Drivers
Adapters for the enterprise service bus, MCP, A2A, channels, and model, speech and storage providers.
One thing: a language effect is a kernel step, and that step is the span, the event and the checkpoint.
Why it is an operating system, not a list of features.
The effect is the step
Writing call or decide creates a durable step, and replay and trail come with it.
The actor is the agent
Subagents are child actors, and stopping an agent stops the whole tree.
Policy applies at every edge
One declared aspect governs gateway, tools, input, voice and memory.
Operating agents is the work left over after deployment.
What keeps an agent fleet up in production, from what it does to what it costs, and from the first run to the thousandth.
Agents introduce a new operational unit. Tessera treats agents, runs, prompts, models, tools, policies, tokens and decisions as first-class entities, not as lines scattered across service logs.
Inspect runs, prompts, models and tools, side by side.
Measure latency and failures per agent, not per service, along with the cause.
Investigate behavior without stitching together disconnected logs from different systems.
Treat runs and prompts as entities, not services, each with its own history.
Understand how agent workloads consume models and infrastructure across the organization. Track usage and cost at the level your business actually operates, and find out early which agent costs more than it delivers.
Track consumption by agent, team and department, throughout the month.
Compare cost across models, workflows and environments, on the same scale.
Define limits before the invoice becomes a month-end surprise.
Attribute every dollar to the team that generated the spend, with no arbitrary allocation.
An enterprise AgentOS sits on the critical path of production workloads. In the Tessera Runtime, every run is an actor in a cluster, with a checkpoint at every step: it resumes where it stopped, identically, and parallelism grows without concurrency code.
Run concurrent workloads with efficient resource usage, even at peak.
Promote revisions across environments under the same policy, from test to production.
Roll back to the previous revision without losing the trail of what happened.
Scale parallelism without multiplying infrastructure at the same rate.
Built for environments where control isn't optional.
Tessera was designed for organizations where security, governance, traceability and compliance are production requirements, not review steps at the end of a project. The controls below apply to every agent running under Tessera, whatever the team that built it.
Policy enforcement
Rules defined at the organization level and enforced on every agent that runs.
Access control
Which agents can reach which systems, data and tools, and under which identity.
Human approval
Critical actions pause before they execute and wait for an authorized person.
Audit trails
A record of what each run did, with the configuration that was active at the time.
Environment isolation
Development, staging and production kept apart by credentials and by policy.
Data governance
How sensitive data is masked, retained and exposed to external models.
Change traceability
Who changed a prompt, policy or version, when, and what the change caused.
Security controls
Secrets, credentials and keys kept out of the code and out of the agent's reach.