The short answer: AI agents need protocols, not longer conversations
Microsoft Research is exploring a problem that every serious enterprise AI program will soon face: once you move from one model answering one prompt to many agents coordinating work, natural language becomes expensive, noisy, and hard to govern.
The research direction, known as OPTiMACS, asks a practical question: should AI agents always communicate in ordinary sentences, or can they learn more efficient message formats for different tasks?
That question may sound technical. It is not. It touches enterprise architecture, operating models, finance, risk, and the future role of IT.
The next generation of AI agents will not be judged only by how well they reason. They will be judged by how efficiently they coordinate.
Why natural language is a weak operating layer for agentic systems
Large language models are excellent at producing fluent text. That does not mean fluent text is the best way for machines to coordinate complex work.
Human language carries ambiguity. It is verbose. It repeats context. It can hide assumptions. It is also costly because every token has a price, every long exchange increases latency, and every additional paragraph creates more surface area for misunderstanding.
In a single chatbot session, this may be tolerable. In a multi-agent workflow, it becomes operational debt.
Consider an enterprise scenario where agents are responsible for triaging customer requests, checking policy constraints, retrieving account data, drafting a response, escalating edge cases, and logging evidence. If each agent explains its full reasoning in prose to every other agent, the system may appear intelligent but behave inefficiently.
The outcome is familiar to anyone who has managed large operational processes:
- Too much internal chatter
- Higher cloud and model costs
- Slower cycle times
- Harder auditability
- More opportunities for misinterpretation
- Less predictable performance at scale
This is why agent communication is not a side issue. It is one of the foundations of production-grade AI.
What OPTiMACS gets right
The interesting part of Microsoft’s research is not merely that agents can use structured messages. That idea already exists in APIs, schemas, function calls, event streams, and workflow engines.
The more important idea is that communication itself can become adaptive.
OPTiMACS treats the communication format as something the system can optimize, rather than a fixed design decision made once by developers. The research frames this through an Expanding Markov Decision Process, meaning the system can explore and expand the set of possible message representations over time.
In business language, that means the agents are not limited to either free-form conversation or rigid templates. They can move toward a middle layer: structured enough to reduce ambiguity, flexible enough to fit the task.
That middle layer is where enterprise value will likely emerge.
The enterprise lesson: less language, more control
Many organizations are still impressed when two agents hold a long conversation. They should be more impressed when two agents complete a task with a compact, auditable, low-cost exchange.
A mature agentic architecture should separate three communication layers:
- Machine-to-machine coordination for efficient execution
- Human-readable logging for supervision and audit
- Management-level reporting for performance and risk oversight
These layers should not be confused. Agents may need compact internal signals to work efficiently, while humans need clear explanations to govern outcomes. If we force every internal agent exchange to look like a human memo, we pay a performance tax. If we allow agents to communicate in opaque formats without explanation layers, we create a governance problem.
The right answer is not natural language everywhere. The right answer is controlled translation between machine efficiency and human accountability.
The finance angle: token waste becomes margin leakage
In early AI pilots, token cost is often treated as a minor line item. At scale, it becomes a meaningful operational metric.
A multi-agent process can multiply usage quickly. Five agents exchanging ten long messages per case across thousands of daily cases can turn an elegant prototype into an expensive production system. Add retries, monitoring, tool calls, retrieval, and audit logs, and the economics become even more sensitive.
Better agent communication can improve the financial model in several ways:
- Lower token consumption
- Reduced latency per transaction
- Fewer failed or repeated steps
- More predictable infrastructure costs
- Better capacity planning
- Easier measurement of process efficiency
This is where AI strategy must connect to finance. The question is not whether an agent can perform a task in a demo. The question is whether the process can run reliably, economically, and safely at enterprise volume.
Human in the loop still matters, but not everywhere
AI enables organizations to automate non-deterministic work, the kind of work that previously required judgment, interpretation, and contextual decision-making. That is precisely why human oversight remains critical.
But there is a common mistake: adding a human checkpoint to every step and calling it responsible AI.
That approach does not scale. If every AI-assisted process requires the same amount of human attention as before, the organization has not redesigned the operating model. It has only added software to the old process.
The better model is supervisory leverage. A person who previously handled one process should be able to supervise hundreds of AI-executed processes through exception management, sampling, dashboards, and escalation rules.
In practical terms, human-in-the-loop design should focus on:
- High-risk decisions
- Low-confidence outputs
- Regulatory exceptions
- Material financial impact
- Customer harm scenarios
- Unusual behavior patterns
The goal is not to remove people. The goal is to move people from repetitive execution into scalable supervision.
Why this is not just an IT problem
AI agents sit at the intersection of technology, business process, compliance, human behavior, and management. Treating them as a purely technical implementation is one of the fastest ways to create unstable systems.
A strong enterprise AI program requires deep knowledge in several areas:
- AI architecture and model behavior
- Business process design
- Data governance
- Risk and compliance
- Change management
- Operational finance
- Domain expertise
This is also why academic research matters. The most valuable AI work is increasingly multidisciplinary. It is not only computer science in isolation. It is the study of how intelligent systems interact with real workflows, human decisions, organizational incentives, and measurable outcomes.
There are many self-appointed AI experts in the market. Large enterprises can usually filter them out. Small and mid-sized businesses are more exposed. Poor advice in AI is not harmless; it can lead to wasted budgets, weak security, broken processes, and unrealistic expectations.
Education, real business experience, and applied implementation knowledge are not optional in this field. They are the difference between experimentation and operational capability.
What organizations should do now
Microsoft’s research is not a product checklist, but it points to decisions organizations should start making immediately.
First, companies need an internal capability for building and managing AI agents. Buying tools is not enough. The organization needs standards, patterns, reusable components, monitoring, governance, and talent.
Second, enterprises should advance on two tracks at the same time: AI literacy for employees and agent development as an organizational capability. These are different adoption paths. AI tools often require employees to change their work habits. Agents, when designed well, can automate or orchestrate processes with less day-to-day behavior change for end users.
Third, agent platforms matter. Microsoft Copilot Studio is a reasonable option for organizations invested in the Microsoft ecosystem, and it is improving. At the same time, tools such as n8n are entering environments that once would have rejected them as too lightweight for large enterprises. The market is becoming more open, more modular, and more competitive.
Fourth, vendor choice should be pragmatic. Claude remains one of the most effective systems for broad enterprise use and applied work, especially with tools such as Claude Code and collaborative workflows, although security and data controls must be evaluated carefully. Microsoft Copilot is becoming more capable, even if large-platform innovation can feel slower than faster-moving AI-native companies. OpenAI continues to offer strong and varied foundation models. The right enterprise strategy is rarely blind loyalty to one vendor.
IT will become HR for agents
As agent adoption grows, IT departments will take on a role that looks surprisingly similar to human resources.
They will need to know which agents exist, what they are allowed to do, who manages them, how they are evaluated, when they should be retired, and how their performance is reviewed. Agents will need onboarding, permissions, monitoring, incident handling, and lifecycle management.
This is not science fiction. It is the natural result of giving software systems more delegated responsibility.
A future-ready organization should be able to answer simple questions:
- Which agents are active in the business?
- What processes do they influence?
- What tools and data can they access?
- What performance metrics define success?
- Who is accountable when an agent fails?
- What communication formats do agents use internally?
- Which decisions require human escalation?
If those questions cannot be answered, the company does not yet have an agent strategy. It has a collection of experiments.
The real breakthrough is operational discipline
The most important message from Microsoft’s work is not that agents may invent better ways to talk. It is that communication design will become a competitive advantage.
Enterprises that let agents chatter endlessly in natural language will face cost, latency, and governance issues. Enterprises that over-constrain agents with rigid workflows will limit their usefulness. The winners will build adaptive systems with clear boundaries, measurable performance, and human oversight where it actually matters.
The next phase of AI will not be won by organizations that simply deploy more models. It will be won by organizations that understand how to turn models into managed, efficient, auditable business processes.
AI agents do not just need better answers. They need better coordination. And organizations need the maturity to manage that coordination before it becomes the next hidden bottleneck.