A single API call can now kick off a coordinated AI workflow, with multiple agents and models working behind the scenes. That’s powerful, but it also creates a new blind spot for enterprise leaders: activity, cost, and risk are moving faster than visibility.
A multi-agent workflow chains several agents together, and each is specialized for a different part of a process. One researches, another drafts, another reviews, and another routes the result. They may run sequentially or in parallel, with context passed automatically through an orchestration layer.
What makes this significant for enterprise leaders is tools like Sakana Fugu, which Sakana describes as a system that dynamically orchestrates a pool of models and expert agents through one OpenAI-compatible API. From the enterprise’s perspective, that can look like a simple API call. Inside, the platform may be selecting models, assigning roles, coordinating handoffs, verifying intermediate work, and returning one final answer.
Sakana also says the specific models Fugu selects and how it coordinates them are proprietary and not exposed by design. For CIOs, that’s the visibility gap: orchestration can work correctly, but the organization won’t have a step-level record of agent activity across teams, workflows, and controls.
When a single API call obscures the internal steps of a multi-agent workflow, enterprise monitoring may show the request, response, and request-level cost. It may not show which agents ran, which models were called, which retrieval or tool calls happened, how many tokens each step consumed, where errors were corrected or propagated, or who owns the downstream decision.
Gartner predicts that up to 40% of enterprise applications will include task-specific agents by 2026, up from less than 5% in 2025. Gartner separately predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. That’s the problem: adoption is accelerating, while the governance infrastructure to measure and control agentic work is still catching up.
Traditional monitoring asks: Is the system up? Agent observability asks: What did the agent do, why did it do it, what did it cost, and was the output right?
Standard application performance monitoring tools and log aggregators can show latency, errors, uptime, and infrastructure events. They usually don’t capture the decision path across prompts, retrieval steps, tool calls, model handoffs, and evaluation checks. That’s the gap between monitoring software behavior and governing agent behavior.
The failure mode is different too. Traditional software usually fails in a way an alert can catch. Agents can drift, loop, over-retrieve, choose the wrong tool, or pass a weak answer to a downstream agent. A run may complete successfully, but it can still produce a compliance problem, wrong customer outcome, or expensive retries.
In our work with clients, the organizations most exposed to multi-agent cost overruns are the ones that deployed agents quickly and added observability later. By the time the budget problem surfaces, months of unattributed spend have already accumulated.
Useful visibility has to follow the work, not just the endpoint. That means three capabilities.
Effective observability requires capturing each agent’s inputs, outputs, tool calls, model invocations, retrieval steps, handoffs, and final answer. Each run should have a trace ID that connects the original request to every intermediate step. Without that chain, teams debug by screenshot and anecdote.
Multi-agent workflows multiply token consumption because each agent may call a model, retrieve context, invoke tools, or retry steps. A retrieval agent that runs fine at 1,000 calls per day can become a significant budget problem at scale if nobody’s monitoring cost at the step level. Attribution connects spend to the specific workflow, team, and tool generating it.
When a team uses Fugu, CrewAI, or another orchestration platform, internal agent activity may be hidden by a single endpoint. Visibility requires either a direct data integration with that platform or a gateway in front of it. The pattern to watch is if agent tooling gets easier for every team to deploy, while central visibility gets harder unless it is designed in up front.
Larridin’s Token Spend & Insights captures AI spend generated by people, tools, and automated workflows, then attributes token costs to workflows, teams, and platforms so agent-driven activity does not disappear inside generic API usage.
For CIOs, multi-agent visibility is part of the broader shift from AI experimentation to AI measurement and optimization. Larridin helps enterprise leaders see where AI is being used, what it costs, which workflows are driving activity, and where governance needs to catch up.
Book a Discovery Call to see how Larridin surfaces multi-agent activity across your organization.
A multi-agent workflow is a process where multiple AI agents collaborate on a task, each handles a specialized portion, and they’re coordinated by an orchestration layer. Agents can run in sequence or in parallel, passing context between them automatically, with no human initiating each individual step.
Single agents are easier to trace because the path from request to response is shorter. Multi-agent systems multiply the failure surface because errors in one agent can propagate silently through downstream agents. Orchestration platforms that hide multiple agents behind a single API make this harder: you see the input and output, but not the steps in between.
Each agent in a multi-agent workflow may make its own model calls. One outside request may involve multiple prompts, retrieval steps, tool calls, or retries. Token costs compound with each step, and without step-level attribution, teams may know the bill went up but not which workflow created the increase.
Effective multi-agent observability should trace every step, not just endpoints. Look for hierarchical trace capture that preserves parent-child relationships across agent handoffs, token cost attribution at the workflow and team level, integration with orchestration platforms you do not control, and consistent behavior across cloud and model environments. Vendor lock-in at the observability layer is particularly expensive because historical trace data carries long-term operational value.
Not sure how much of your AI spend is coming from agents versus people?
Book a Discovery Call and get full visibility into your multi-agent workflows in days.