[agenticwork]
← blog

architecture · july 8, 2025 · 9 min read

Multi-Agent Orchestration: How the Platform Coordinates Specialized AI Agents

A single AI agent talking to a single model is a chatbot. An enterprise AI platform needs multiple specialized agents working together: one for conversational interactions, one for code generation, one for deep research, and one for multi-step workflow execution. The hard part is not building the agents. The hard part is governing them.

Our platform runs four specialized agent types, each optimized for a different class of work, all governed by the same security, audit, and approval infrastructure.

The Agent Types

ChatMode: Conversational AI with Memory

ChatMode is the conversational interface. It handles multi-turn dialogues with a three-layer memory system:

  • Session memory: Active conversation context with automatic summarization when context windows fill. The agent remembers what you discussed this session and can reference earlier points.
  • User memory: Persistent profile data — preferences, prior decisions, project context. Accumulated across sessions so the agent improves over time.
  • Global memory: Organization-wide knowledge — engineering standards, compliance requirements, approved vendors, team structure. Shared across all users so organizational knowledge is not siloed.

ChatMode integrates with RAG (Retrieval-Augmented Generation) for context injection from knowledge bases, document stores, and vector databases. The agent does not just respond from model training data — it pulls relevant context from your organization's actual documents and systems.

CodeMode: AI Pair Programming in the Browser

CodeMode provides VS Code in the browser with AI pair programming capabilities. Code generation, refactoring, and execution happen in K8s-isolated sandboxes — the generated code cannot access production systems, cannot modify files outside its workspace, and cannot persist beyond the session.

CodeMode is particularly useful for MCP server development. Teams build and test MCP servers directly in the platform, with sandbox isolation ensuring that a buggy MCP server cannot affect the cluster. When the server is ready, it deploys through the standard CI/CD pipeline with the platform's governance controls applied automatically.

Research Mode: Autonomous Multi-Phase Investigation

Research Mode is an autonomous research pipeline for complex questions that require gathering information across multiple sources, synthesizing findings, and building structured deliverables. The pipeline handles problem decomposition, source retrieval, cross-source synthesis, finding validation, report generation, and citation attribution.

Research Mode builds on context accumulated across sessions. A research project started on Monday can continue Thursday with the agent retaining the accumulated findings, identified gaps, and synthesis progress.

AgenticWorkflows: Visual Multi-Step Orchestration

AgenticWorkflows is the Workflow Builder — a visual interface for constructing agent pipelines. Each node in a workflow is a distinct operation: data extraction, API call, model inference, human approval, conditional branching, or result aggregation. Workflows can compose the other agent types — a workflow step might invoke ChatMode for user interaction, CodeMode for code generation, or Research Mode for investigation.

Workflows support HITL approval gates at any node. A data extraction step can auto-approve (low risk, read-only). A database write step requires human review. A deletion step requires escalated approval. The workflow pauses at gates and resumes when approved.

Unified Governance

The key architectural decision is that all four agent types share the same governance infrastructure:

  • Same Audit System: Every action by every agent type is recorded in the same immutable, cryptographically hashed audit trail. A ChatMode conversation and a CodeMode execution appear in the same audit log, searchable by user, time, agent type, or action.
  • Same credential isolation: Whether a tool is invoked from ChatMode or AgenticWorkflows, it gets the same scoped, time-limited credential. The credential isolation layer does not care which agent type initiated the request.
  • Same RBAC: User permissions apply across all agent types. If a user does not have permission to invoke the AWS capability, they cannot invoke it from ChatMode, CodeMode, or a workflow.
  • Same SmartModelRouter: Model selection is centralized. ChatMode and AgenticWorkflows use the same routing logic, the same Intelligence Slider settings, and the same failover chains.
  • Same DLP scanning: Sensitive data detection runs on all model requests regardless of which agent type generated them.

Agent Coordination Patterns

Multi-agent orchestration on the platform follows a few key patterns:

Sequential Handoff

A workflow step calls Research Mode to investigate a topic, passes the findings to ChatMode for user discussion, and then sends the refined requirements to CodeMode for implementation. Each handoff passes structured context, not raw conversation history. The receiving agent gets what it needs without inheriting the full context of the previous agent.

Parallel Investigation

A workflow fans out to multiple Research Mode agents simultaneously — one investigating market data, one analyzing competitor products, one reviewing internal documentation. Results are aggregated at a merge node. Each parallel agent operates in its own sandbox with its own credentials.

Human-Gated Escalation

An agent hits a decision point that exceeds its authorization. Instead of making the decision, it creates a structured approval request and pauses. The request routes to the appropriate human reviewer based on RBAC policies. The human approves, modifies, or rejects. The agent continues with the decision.

Why Not One General Agent?

A single general-purpose agent that handles conversation, code, research, and orchestration sounds simpler. In practice, it is worse. Specialized agents can:

  • Use different models for different tasks (economy for extraction, premium for reasoning)
  • Maintain specialized memory structures (session vs research accumulation)
  • Apply different sandbox configurations (browser-based for code, network-restricted for tools)
  • Scale independently (more CodeMode capacity during development sprints)

The governance layer unifies them. The specialization keeps them efficient. This is our platform's approach to multi-agent orchestration: specialized agents, unified governance, shared security infrastructure.