AgentPlat
AgentPlat gives your application the foundation to coordinate teams of agents, preserve shared work across sessions and interruptions, and keep people in control of critical decisions—all on your infrastructure.
npm view @agentplat/framework dist-tags
New integration · source preview
Developers who choose TypeSafe Jev can turn assessment output into typed signals for classifying or routing work. Your application still selects the data and rubric, and keeps control of policy, human approval and protected actions. The adapter is implemented in source, is not on npm yet, and has no live quality or cost comparison.
See what the integration is for → · Read the Jev assessor docs
For developers
Build applications where humans and agents work together, with persistent state, approvals and controlled execution on your own infrastructure. Reuse the work model around the model call.
Agent Rooms connect participants, messages, tasks and versioned artifacts, so you can reuse a collaboration model instead of designing one from scratch.
Model approvals, contributions and interventions as part of the work. A person can add context or request changes in a process that outlives a session.
Use task dependencies, capability-based assignment and typed handoffs to make who does what, and what happens next, explicit.
Durable storage adapters, revision checks and leases support recovery and controlled retries. Preventing duplicate external effects still depends on your integrations.
Inspect artifact versions, events and agent-definition references to reconstruct the work and give reviewers useful context.
Public interfaces for models, storage, execution, tools and events let you replace infrastructure while keeping your application's work model.
Illustrative workflow
With durable adapters configured, work can persist across human waits and recoverable interruptions. You build the business logic, UI and integrations.
Open source · Apache-2.0
Inspect the implementation, adapt behavior and build your own adapters. Self-host without private AgentPlat services, choose where data lives and keep the code available for maintenance.
Your team operates hosting, inference, identity, backups and connectors. Data handling depends on the providers and configuration you choose.
Developer preview. APIs may change. Review current status and integration details.
A strong fit for TypeScript teams building internal agent platforms,
SaaS workspaces with approvals, or long-running collaborative automation.
For a simple text-generation feature, start small with AgentPlat.ask.
AgentPlat on npm
Bring AgentPlat into your own product with our family of open-source npm packages. Start with the pieces you need and explore more as your ideas grow.
A foundation for building and running AI agents as part of your application.
Give people and agents a shared place to work, keep context and move decisions forward.
Help agents coordinate around shared goals and work together across your application.
Beyond single-agent automation
Adding more agents is not enough. Critical work needs coordination that contains failures, preserves authority and produces evidence that teams can inspect after every decision.
Human-agent collaboration
Agent Rooms bring people, agents, context, tasks, artifacts, approvals and evidence together in one tenant-scoped operational workspace.
The design origin
Agent Rooms originated in the paper Agent Rooms: A Conceptual Framework for Persistent Human-Governed Multi-Agent Collaboration by Douglas Rodriguez. The paper defines the persistent, governed collaboration container behind this product direction.
Read the paper and cite the framework → Explore the research →New research · Preprint v1.0
The paper examines how multi-agent organizations can change at runtime while preserving bounded authority and causal continuity across independently owned subsystems. Its evidence compares governed reconfiguration with fixed, minimally adaptive and durable-workflow baselines, and documents how losing proposals with partial effects are reconciled.
Agent Morphogenesis
AgentPlat Agent Morphogenesis coordinates governed changes to agents, roles, teams and strategies. Recruit a specialist, create an agent from an approved profile or reorganize work within explicit policies and budgets.
Decisions can come from an authorized agent, a person, a collective or local policy. Approval, execution permission and the resulting effects remain separate, inspectable steps.
The operating foundation
Collective resilience is not a library call. It requires a common operating layer for applications, humans, workflows, memory, agents, models and infrastructure.
Product surfaces, customer workflows and internal tools built on top of the platform.
Approvals, escalations, collaboration, review queues and operational handoffs.
Durable orchestration for long-running tasks, retries, branching and policy gates.
Scoped context, state, artifacts and retrieval boundaries for each tenant or process.
Lifecycle, tools, identity, limits, execution state and deterministic audit capture.
Provider-neutral model access across inference, embeddings, routing and evaluation.
Cloud services, storage, queues, secrets, observability and deployment primitives.
Integration responsibilities
The hard part is not adding agents. It is keeping their coordination, authority, recovery and protected actions reliable when conditions change.
The capabilities behind the promise
AgentPlat brings coordination, control and evidence together with the platform components product teams need to operate them.
Bounded local collaboration, shared state and collective planning.
Assignment, membership and communication recovery after failures.
Scoped context, artifacts, retrieval and tenant boundaries.
Protected effects bound to identity, authority and current assignment.
Standardized access to tools, services and external systems.
Traceable decisions, recoveries, tool calls, approvals and outcomes.
Durable orchestration for complex business processes.
Review queues, escalations, approvals and shared accountability.
Users, tenants, agents, roles, permissions and execution scopes.
Metrics, traces, quality signals and collective runtime health.
Environment-aware packaging for platform and product teams.
Limits, approvals, data rules and operational guardrails.
Success stories
Different products. Different ways to turn AI into something people can use. Discover the AgentPlat contribution behind each experience.
HealthcareAgentPlat brings execution controls and traceability to AI-assisted patient profiles and transplant-center matching.
Read the story
AI simulationsAgentPlat powers the agent execution behind Simultrix’s simulated focus groups and bilateral negotiations.
Read the story
Legal technologyBuilt on AgentPlat’s application platform, No Firmes connects document intake, specialized AI analysis and report delivery.
Read the story
Creator marketingPooshlo combines AgentPlat-powered matching and product discovery with profile audits and content collection in an operating system for creator marketing.
Read the story
Media · Daily BriefGrishen used AgentPlat as the design foundation for Daily Brief, turning El País’s reporting into AI-generated summaries for different audiences.
Read the story
E-commerce · Agent RoomsAfter using AgentPlat directly in 2025, Trafilea began developing its own implementation in 2026, drawing on Agent Rooms and targeting an 80% agentic e-commerce operation.
Read the storyUsed By
AgentPlat has supported enterprise workflows, vertical products and customer-facing platforms across finance, commerce, healthcare and legal operations.
Architecture
Use AgentPlat as the shared layer between agents, application teams and infrastructure. Coordination stays local and resilient while operational controls remain consistent.
git clone https://github.com/Agentplat/agentplat.git cd agentplat/examples/rooms-api cp .env.example .env docker compose up --build -d --wait node scripts/proposal-demo.mjs # Requires Git, Node.js 22.13+ and Docker Compose v2. # Scripted review; no live model calls by default.
Architecture and evidence
AgentPlat distinguishes architectural capacity from demonstrated results, so product teams can evaluate each claim precisely.
Control at every boundary
Agentic systems need the same seriousness as any other critical infrastructure: identity, authority, fault tolerance, auditability and operational ownership from the beginning.
Where collective operations matter
AgentPlat is built for operations where multiple specialized agents must coordinate safely across systems, teams and changing conditions.
Coordinate specialized agents across evolving incidents while preserving authority and evidence.
Operate long-running, multi-system work where peers can fail and priorities can change.
Run governed collective automation across existing tools without losing control or traceability.
Principles
AgentPlat is an operational foundation for dependable collective operations—not a prompt demo. It gives teams the structure to build reliable multi-agent systems.
Evaluate AgentPlat
Start with one workflow, verify its behavior and review the integration obligations for your application.
Run a two-agent proposal, request a revision and inspect the retained artifacts and approval.
Run the guide →Distinguish artifact review from protected execution. Check a provider checkpoint and local effect receipt.
Check approval controls →Stop a worker after its result commits and verify that a successor reuses completed work.
Run the recovery scenario →Compare Room collaboration with graph orchestration and step-based workflows.
When to use AgentPlat →Separate source behavior, package distribution, research and observed operational results.
Inspect the evidence →Project-requested model commentary is not an independent evaluation, vendor endorsement or production validation.
Read the archived commentary →Articles
Practical guides to shared work, human decisions and process recovery. Explore all articles →
Connect tasks, versioned artifacts and review decisions around a shared objective.
Read the collaboration guide →Understand the difference between accepting an artifact and authorizing an external action.
Read the approval guide →Learn how stable identities and retained results help a successor reuse completed work.
Read the recovery guide →Contact
Share what you are building, the platform constraints you have and the kind of agentic system you want to operate.
Start Building
Use AgentPlat to coordinate autonomous agents safely, recover from failure and verify every critical outcome.
New research · Preprint v1.0
Read the new preprint on governed runtime organizational change, its comparative mission evaluation and the evidence limits around partial effects and superseded proposals.
