Multiagent systems for business process orchestration

A single AI agent handles a task with a single context window, tool set, and reasoning loop. A multiagent system deploys multiple AI agents with different roles, tool access, and specialisations that collaborate to complete a task. This enables: division of labour across specialised agents (a research agent that searches, an analysis agent that reasons, an integration agent that executes API calls); parallel processing of independent subtasks; context management across long workflows that exceed a single agent's context window; and specialisation that improves per-task accuracy compared to a generalist single agent.

LangGraph is a framework from LangChain that models agent workflows as directed graphs — nodes are agents or processing steps, edges are transitions triggered by conditions or agent outputs. This graph-based approach is popular because it makes complex conditional workflows (if the agent discovers X, go to Y; if an error occurs, route to the recovery node) explicit and debuggable. LangGraph supports stateful workflows (persisting state across steps), human-in-the-loop interrupts, and streaming of intermediate results. It is widely deployed in enterprise multiagent systems and integrates with LangSmith for observability.

Prompt injection is an attack where malicious instructions are embedded in content that an AI agent processes — emails, web pages, documents, database records. When the agent reads this content, it may inadvertently execute the embedded instructions rather than following its original system prompt. In multiagent systems, this risk is amplified: a compromised agent can propagate malicious instructions to downstream agents. Mitigate prompt injection with explicit instructions not to follow instructions found in processed content, sandboxed execution environments, output validation before action agents execute irreversible operations, and human review gates for high-value autonomous actions.

Financial services leads in enterprise multiagent adoption for credit processing, claims management, KYC onboarding, and fraud investigation — processes that combine structured data processing with complex decision making and regulatory requirements. Insurance is a major adopter for claims automation. Healthcare is implementing multiagent systems for prior authorisation, billing, and clinical documentation. Legal services uses agents for contract review and due diligence. IT operations has deployed multiagent systems for incident response and runbook execution. The common thread is processes that combine multiple data sources, conditional decision making, and cross-system actions — exactly where multiagent systems outperform traditional rule-based automation.

Robust multiagent systems handle failures at multiple levels: individual agent retries (retry failed tool calls up to N times with exponential backoff); fallback paths (if agent A fails, route to agent B with a simpler approach or human review); workflow checkpointing (save state at each completed step so failed workflows can resume from the last checkpoint rather than restarting); human escalation agents that receive context from failed workflows and route to appropriate human reviewers; and dead-letter queues for workflows that exceed maximum retries. Observability tooling (LangSmith, Langfuse) is essential for diagnosing failure patterns across large volumes of workflow executions.

Human-in-the-loop (HITL) is an architecture pattern where AI agents pause and request human input or approval for specific steps — typically high-value, irreversible, or ambiguous decisions. The agent prepares context (what it knows, what decision is needed, what the options are), presents it to a human via email, Slack, or a review UI, and waits for the human's input before continuing the workflow. After the human responds, the agent reintegrates the decision and continues. HITL is essential for: financial approvals above defined thresholds, regulatory submissions, customer-facing communications, and any action that cannot be undone if the AI makes an error.

ROI measurement for multiagent BPO covers: time savings (total manual handling time per process × volume × hourly cost, compared before and after deployment); error rate reduction (rework and correction costs reduced by higher accuracy automated processing); throughput improvement (volume processed per time period with unlimited scaling vs human capacity ceiling); exception rate reduction (percentage of cases requiring human intervention, which should decrease as agents improve); and cycle time reduction (end-to-end process duration from intake to completion). Establish baseline measurements before deployment and track for 90 days post-deployment to produce a credible ROI case for programme expansion.

The main costs are LLM API calls (each agent action typically requires one or more model inference calls — multiply by workflow volume and average agent turns per workflow to estimate), infrastructure for orchestration and tool execution, integration development and maintenance for system connectors, observability tooling (LangSmith, Langfuse), and human review capacity for HITL steps and exception handling. LLM API costs are often the largest variable cost — use cost per workflow as the key unit metric and set per-workflow cost budgets to prevent runaway spending. Smaller, fine-tuned models for specific agent roles (vs frontier models for everything) can significantly reduce cost at scale.