Agent Platform은 AI Agent를 만들고, 실행하고, 관리하는 모든 과정을 하나로 연결하는 플랫폼입니다. 사용자는 채널에서 여러 에이전트와 대화하며 작업을 진행할 수 있고, 프롬프트만으로 새로운 에이전트를 생성할 수 있습니다. 개발, 배포, 운영, 모니터링 기능을 통합 제공하여 에이전트 활용 경험을 더욱 쉽고 효율적으로 만듭니다.
Digital Workplace Lablup AI Agent Platform Core System Build fast, Run many, Govern centrally On premises Lablup AI Agent Platform Surface Connection layers of AI Agent Platform
agent use at work Large enterprises: scaling agents 70% Adoption stage Q1 2026 70% 59% 41% Testing or not started 40% 35% 31% 35% Using in real work 27% Running many agents 0% 32% 27% 0% 2025 2026 Stack Overflow. Surveyed technologists. Nearly 2 in one year. 2025 2026 McKinsey. 1B revenue. Others stayed flat at 22 . KPMG. 59 have moved past testing. 27 many agents together. People already use agents, large enterprises are scaling them, and one in four runs many agents together. run
Digital Workforces • Redesigned Work: Agents do the work across documents, data, and systems. People set goals and decide • Agents Multiply: Every repeated task becomes an agent, and agents call each other to finish the job • Scaling Limit: Each agent needs a runtime, security, and oversight. Built team by team, effort and risk grow with every agent • Built Once: Build runtime, security, and oversight once, and every agent uses t hem from day one • Security First: The data agents need most is the data a company must protect most, so agents run where it is kept
BUILD Build in Minutes Create and deploy agents across models, frameworks, and workflows Continuous Agent Evaluation Turn real world signals into measurable improvement OPERATE Production Ready Runtime for Multi Agent Run many agents with durable state, memory, sandboxing, and observability concurrently and coordinate them across channels Shared Agent Ecosystem Share tools, SKILLs, memory, kn owledge, and policies GOVERN Unified Control Plane Manage agents, teams, versions, resources, quality, and costs Governed Autonomy: Enforce identity, permissions, guardrails, approvals, isolation, and auditability by default
evaluated. Three parts: Agent Framework stru cture , Agent Build create and evaluate , Tool Build tools . Agent Builder Agent Framework Agent Discovery Agent Evaluation Create agents from natural la nguage, refine with preview and feedback, run on the Managed Harness External SDK integration, sta ndard structure and behavior , build your Custom Harness Find agents in the registry by name, role, or tools; see publi sh, approve, and deprecate status Measure goal success and trajectory from user and HITL signals, message quality, task outcomes, and runtime signals Tool Builder Agent DevTools Tool Discovery Tool Evaluation Turn REST, Lambda, and MC P servers into tools; docume nt spec and usage Build and test your harness in your Local Environment, then ship it as is Find tools in the registry by s emantic search; load only th e definitions needed into con text Verify tool selection, parame ters and results, failure rate, and efficiency
Agent Builder Agent: Built in agent, create agents from natural language, refine with preview and feedback, run on the Maintained Harness • Maintained Harness: Automatically composed and tool complete, with no user editing required. Ready to run on the platform • Evolutions Prompt: Users can improve their agent as development loop with Human Feedback
eve/AI SDK Vercel , strands Amazon , langchain/langgraph, OpenAI Agents SDK, or your preferred SDK • Custom Harness: Write your own loop, stop conditions, and hooks and run on the same Runtime and Governance
dev environment, run, debug, and deploy your Custom Harness • Local Dev Environment: Start the local development server and terminal UI • Deployment: Deploy local projects
to context HARNESS tools · files memory · skills instruction ACT The harness executes and returns the result REASON Read the current context MODEL PLAN Break the goal into executable steps GOAL
Runs multiple agent loops in parallel and coordinates their work. • Agent Server: Hosts, scales, and manages agents built with different frameworks. AGENT RUNTIME AGENT LOOPS AGENT • Durable Runtime: Keeps sessions and state across long running tasks, failures, and restarts. • Sandbox: Safely runs agent generated code and files in isolated environments. • Gateway: Provides secure, governed access to models, tools, credentials, and other agents. EXECUTION ENVIRONMENT GATEWAY Agent Server Model · Tool Gateway Durable Runtime Credential Broker Sandbox Agent Communication
execution • Stores conversation history and working state • Return to or continue an earl ier session • Isolated per user Memory Bank Artifact Store • Long term memory that outli ves sessions • Files and outputs the agent produced • Facts are extracted when a • Persist after the session end s • session is saved • User corrections and operati ng rules accumulate • Referenced in later decisions • Roll back to earlier versions • Passed as input to other age nts
All are called through the Tool / MCP Gateway and fol low the user s permissions and policy. Browser Tool Grounding Service Document Search Data Connectors Navigate pages, click and t ype, take screenshots. Han dles logged in sites and dy namic pages like human be havior Web search and page fetch . Attaches sources and con denses results to fit the con text Find relevant passages in u ploaded, indexed documen ts. Chunking, embeddings, hybrid search, reranking Query databases, internal SaaS, maps, and third part y APIs. Read only scope Code Interpreter MCP Connector File Run code in the sandbox. Calculations, data analysis, charts, file conversion Register tools from external MCP servers and call them through the same interface Read, write, and edit files; h andle PDF and Office. Outp uts go to the Artifact Store
agent works with, like a person using a browser: search, read, judge, click, type. The model decides; tools act • Tools for AX: Every tool a person uses at work, an agent must use too. Browser, files, terminal, email, internal apps • Meta Tool: Default tools to create advanced tool for real world problems. The agents will combine or update these to create tools • Tool Builder Agent: Built in agent that creates new tools from natural language by combining the default tools
the way a person does: navigate, click, type, screenshot, on logged in and dynamic pages • Multi modal: Not only text, extract and analyze multi media which make agents generate much better response
so users reach only the agents they are allowed to • Agent identity: each agent gets its own ID • Outbound: access Slack, GitHub, M365 on behalf of a user or autonomously OAuth, API keys • Virtual Key: a logical key that hides real credentials and binds user, org, project, mod els, and policies
Session stdout, agent actions and events Live status, alerts Runtime Agent Server · Sandbox Broker · Ag ent Harness · Durable Runtime · Age nt Communication Audit, replay, root cause analysis Logs LLM conversations, transactions, tool call inputs and outputs Metrics Goal success rate, tool selection accuracy, guar drail violations, tokens and cost, latency Trace Reasoning tool selection call elegation paths between agents Audit Who used which agent with which credentials an d tools Feedback User ratings and corrections; HITL approve, edit, reject; reviewer comments Gateway Model Gateway · Tool / MCP Gatewa y · Credential Broker Governance Identity · Guardrail · HITL Service · Channel Dashboards, budget control, quality trends Locating where a wrong decision started synthesis; d Accountability, approval history, compliance Input to Evaluation · Optimization Observability API OTel
again 1 Collect Production traces, user f eedback, and HITL decis ions pile up as the evalu ation dataset Human as judge 2 Evaluate 3 Diagnose Score every run automat ically: goal success, tool use, safety, cost Group failures into clust ers. What is broken and where it started the loop never stops 4 Improve 5 Release Get ranked suggestions for instructions, tools, a nd policies. A person ap proves what to apply Re run against the golde n dataset, compare versi ons, roll out safely, roll b ack instantly LLM as judge Collected traces become the golden dataset for re validation production data becomes an asset
Agent: Built in agent that turns traces and feedback into training data, then tunes, tests, and serves the model in one FastTrack pipeline. • Agent to Model Pipeline: It improves the whole agent, not just the model: prompts, tools, and the model get better together. • Right Resources per Step: Agents get tuned models without waiting on infrastructure teams, and improvements reach production faster. • Reusable Pipelines: Build with drag and drop or CLI, share templates across teams, and serve improved models through the Model Gateway
single point every external call from an agent passes through. Model, tool, credential, and ag ent to agent access are gathered Model Gateway Tool / MCP Gateway Credential Broker Agent-to-Agent Models from many providers be hind one API. Routing by quality or cost via Continuum Router Scattered tools and MCP server s behind one endpoint. Semanti c tool search, scope and approv al enforcement Token vault, OAuth and API keys . On-behalf-of or autonomous access Subagents registered as tools, called by name from the registry (A2A) Policy Enforcement Scope · Approval Quota · Cost Audit · Trace Safety Policy and Guardrails appli ed at call time; block on violation Allowed range per agent; risky ca lls pass only after HITL approval Rate limits, per-team cost attrib ution, model downgrade on bu dget overrun Who used which model, tool, a nd credential; forwarded to Obs ervability
Managed Harness AGENT GATEWAY Policy Enforcement Scope · Approval Quota · Cost Audit · Trace Applied to every call Agent B Custom Harness LLM providers Backend.AI · external Model Gateway Routing by quality or cost via Continuum Router Tool / MCP Gateway One endpoint, semantic tool search Orchestrator Agent to Agent Observability Logs · Metrics · Trace · Audit Credential Broker Token vault, on behalf of access Tools · MCP servers Built in Tools External services Slack · GitHub · M365 Called by name from the registry A2A Other agents
Create, deploy, version, delete agents and tools. The base for CLI an d Console • Builder API: Generate agent and tool specs from natural language, run previews, ap ply feedback • Session · Memory API: Create sessions, talk to agents, resume, read history. Mana ge long term memories • Channel API: Integrate Slack, Teams, and other services. Receive messages, respo nd, subscribe via WebSocket and Webhooks • Sandbox API: Create sandbox instances, run commands, move files, destroy • Model Gateway API: List models, set routing and budgets, read usage • Scheduler / Trigger API: Trigger runs by time, event, or condition. • Evaluation API: Manage test cases and golden datasets, run evaluations, read resul ts, submit feedback.
Workspace: Built on top of the Lablup Agent Platform APIs, Builder, Channel, Agent. What Space does, you can build your own Space too • Channels: People and agents in one channel. Ask, watch the thread, reply • Builder: Build, preview, deploy to a channel • Agents / Tools: Find agents and tools: Search the registry, add to your channel • Observability: Open any run to see the steps, tool calls, and results
Stack: Swap storage, auth, and models for any site • Distributed Runtime: Run on any infrastructure, from a single server to Kubernetes clusters • Agent Workforce: Built in agents for common business tasks • Domain Expert Harness: Built in harnesses tuned to each business domain for more accurate work • Tool Studio: Connect and migrate internal systems to the platform without writing code • Workflow Canvas: See every agent on one board. Link agents and approvals visually • Scope Guard Red Team Agent: Scope Guard checks every call against the task. Red Team Agent attacks every new agent before release • Warden • Outcome Evaluation: From AI usage to agentic AI outcomes, from token maximization to minimization. Score every step and cost, and improve them Intelligent Approval System: Approves safe tasks automatically by risk level
lablup
palkan
kastner
chrislema
baasie
marketingsoph
keithpitt
thirion
danielanewman
lara
lfama
elsirapls
axbom
