How Is This Different From Claude Code Or Codex?
Codex, Claude Code, Cursor, Copilot and Pi can turn an ambitious brief into working software, tackling complex problems for days with little intervention. They are built around delivering a result, with a person setting direction and judging the outcome. Meerkat is built for agents that become part of a running operation. They respond to events, coordinate work and carry responsibilities for months. Tasks come and go; the agents keep their identities, memory and roles. Teams can span machines, with agents communicating directly and coordinating shared work. Your application controls membership, connections and authority. That distinction shapes the runtime itself: native asynchronous execution, critical lifecycle rules checked mathematically, and scheduling, recovery and access control designed into how work proceeds. It also shapes how much of the runtime you need to deploy.From Minimal Automation To A Thousand Cooperating Agents
A small runtime for a small job
For a simple automation problem, executable size, startup time and memory use matter. A minimal Meerkat configuration can produce an executable of around 20 MB, with harness startup measured in milliseconds. Its Rust runtime compiles to native code; you select optional capabilities at build time and compose the models, tools and storage your application needs. A command-line task can userkat; an embedded application can call the runtime directly,
in its own process.
This lets you build a compact agent without bringing along an entire interactive
development environment. Small configurations can target embedded devices such
as the ESP32. WebAssembly runs the harness in a browser. A full native build
provides a capable harness in a data center. These are configurations of the
same foundation, with shared agent and provider contracts.
See composability,
build profiles and
browser/WASM examples.
A running operation with many agents
At the other end of the spectrum are operations with hundreds or thousands of agents: continuously available, communicating and sharing work over months. Each agent has an identity, responsibilities, memory and permissions. Individual tasks finish; the operation continues. Meerkat has reached operational peaks of roughly 1,000 agents, with synthetic benchmarks exercising 10,000 agents. Native asynchronous execution and distributed coordination are fundamental to its design. Meerkat provides the machinery for that operating model:- Native asynchronous execution. Agents respond to events and coordinate while other agents wait on models, tools or external systems. Their lifecycle extends beyond an individual request or model call.
- Direct communication and distributed teams. Agents exchange signed peer messages within a host or across machines. Applications define team membership, placement and who can communicate with whom.
- Continuity across execution. A member keeps its identity when it restarts, reconnects or moves to another host. Durable sessions, schedules and work records preserve the state needed to continue; Meta supplies agent memory across tasks.
- Shared work and recovery. Work can arrive from a schedule, an external event, a person or another agent. Runtime services track execution, completion and recovery as part of the same system.
Engineering The Rules That Keep It Running
Long-running systems encounter races, failures and changes of authority. A tool finishes as cancellation arrives. A worker restarts after accepting a job. A permission changes while delegated work is in flight. The runtime needs precise rules for each case. Meerkat defines critical lifecycle decisions as state machines. Shared definitions generate both TLA+ specifications and the Rust kernels used by the runtime. Model checking verifies safety properties across the reachable states of the configured models. These rules govern session execution, coordination, scheduling, authorization lifecycles and recovery. The result is explicit, mathematically checked behavior around questions such as which transitions are legal, who may commit them and what counts as completion. The machine authority reference connects those rules to their definitions, generated code and verification process. Identity and access control are part of that operating model. The system separates who requested work, which agent executes it, on whose behalf it acts and which account supplies credentials. Configured policy can narrow delegated authority and evaluate the concrete operation; process confinement controls what executed commands can access. Meta adds application policy and approval interfaces, while Recall applies access control to shared knowledge. See authorization for attribute-based policy, delegation and host integration.A Full Platform When You Need It
Composability extends to a complete agent system. Core includes tools, MCP, skills, hooks, structured output, compaction, semantic memory, image generation and live voice alongside execution and coordination. The integrated platform adds:- Meta: agent identities and memory, routing, application policy and a console framework you can adapt to your own product.
- Connect: agent-driven integration discovery and preparation, using its library or researching packages for the job.
- Recall: a knowledge system with source evidence, relationships, access-controlled retrieval and maintained views.