Two very different industries are converging on the same insight: AI agents work best when they're narrow, specialized, and kept on a tight leash with human oversight built in from the start. A telecommunications company managing a 5G network core and Salesforce's legal department have each deployed multi-agent AI systems that reject the temptation to let large language models (LLMs) run wild. Instead, they're using coordinated agent networks with safety policies written as code, classical anomaly detection as a gatekeeper, and humans positioned as decision-makers rather than afterthoughts.

The results suggest a pattern: when you treat agentic AI as a team sport rather than a solo performance, the technology becomes production-ready in high-stakes environments where mistakes carry real consequences.

Why Traditional AI Approaches Fail in High-Stakes Operations?

Both organizations started by rejecting two dominant patterns in the industry. The first approach pipes all security telemetry, asset context, and operator queries into a single large model and asks it to emit detections and responses. This fails for three compounding reasons: the context window is too small for the breadth of real-world data, the model is non-deterministic in ways unacceptable for outputs that trigger automated isolation on production infrastructure, and a single prompt edit changes the behavior of the entire system, making change management nearly impossible.

The second pattern treats AI as a productivity layer on top of existing systems. Most vendors ship an LLM assistant that summarizes alerts and drafts playbooks on top of an existing stack. It delivers analyst productivity gains, but it does not change the structural shape of the operation. Rules are still hand-written, and time-to-coverage on a new threat class is still bound by the speed at which humans can write detection logic.

For a 5G core network generating more security telemetry per hour than most enterprise security operations centers (SOCs) see in a week, this bottleneck becomes critical. The real constraint is not analyst triage; it is the detection-engineering team's ability to keep the rule base aligned with a threat landscape that evolves faster than rules can be written.

What Does a Multi-Agent Architecture Actually Look Like?

The telecommunications company's approach treats the SOC as a multi-agent system in which narrow, specialized agents collaborate through an open coordination protocol, integrate with their environment through an open context protocol, gate expensive LLM inference behind classical anomaly detection, and keep a human in the loop by design rather than by accident.

The architecture rests on several non-negotiable properties:

• Narrow Agent Design: Each agent has a single-page contract that fits on one page, including its system prompt, tool list, output schema, and escalation contract. If you cannot read an agent contract on one page, the agent needs to be decomposed into smaller pieces. This discipline prevents the temptation to give agents rich personalities and elaborate prompts, which becomes technical debt.
• Open Coordination Protocol: Inter-agent traffic uses the Agent-to-Agent (A2A) protocol, which was open-sourced by Google in April 2025 and transferred to the Linux Foundation in June 2025. The useful life of a production multi-agent system is measured in years, so the coordination layer's durability dominates any specific framework's convenience.
• Environment Integration via Model Context Protocol: Every interaction an agent has with the underlying environment goes through Model Context Protocol (MCP) servers. This includes asset inventory lookups, current network topology, recent incident history, and configuration state. When a new asset class is added or an inventory system is replaced, the change is absorbed by an MCP server's adapter; no agent prompt needs to be rewritten.
• Safety as Code, Not Prompts: No agent may initiate an externally visible action without explicit approval from a reviewer agent whose safety constraints are codified, version-controlled, and independently testable. The reviewer calls out to a separate policy layer and acts on a deterministic verdict, rather than having the LLM reason about safety on the fly.
• Anomaly Detection as a Gate: A lightweight Isolation Forest pre-filter on raw telemetry routes the LLM-driven agents only to genuinely novel samples. This approach makes the architecture reactive, where inference cost stays bounded, latency stays predictable, and the LLM does the work only humans were doing before.

The results from the 5G core deployment have been substantial. Measured against a baselined set of incident classes, the architecture has reduced mean times to detect and to respond by approximately 40 percent each, has autonomously generated more than 80 detection rules that would not previously have been written, and has compressed the human work required to produce a new rule from roughly three hours to 15 minutes.

How Is Salesforce Applying the Same Principles to Legal Work?

Salesforce's legal and corporate affairs team is building an AI-native, trust-first agentic legal function that holds the promise of defining the best of human-AI collaboration. The team manages hundreds of thousands of requests per year, and this transformation has fundamentally changed how they operate, govern, and partner with colleagues across the company.

The legal team uses Slackbot, an integrated AI companion, to automate routine workflows and give attorneys time back to focus on building trust, calibrating risk against opportunity, and navigating high-stakes complexity. Because Slackbot leverages curated source-of-truth internal knowledge articles and legal resource libraries built by trusted teams, it delivers accurate, consistent guidance the moment people need it.

Real-world legal use cases include:

• Contract and Document Search: In the past, checking the status of a main service agreement (MSA) or finding a specific business associate agreement (BAA) contract meant switching between multiple Salesforce instances and databases. Now, the team asks questions in plain English directly in Slackbot, like "What's the status of the MSA negotiation with [Company X]?" and receives a direct link to the record with no context-switching.
• Streamlined Verification and Intake: The legal team has built a product legal front door in Slack, a self-serve Slackbot-powered intake experience that helps technology and product colleagues get to the right legal resource quickly. Depending on the query, Slackbot can route users to existing legal channels, prompt them to submit a customer case, determine whether a product legal review is needed for a new feature, or walk them through an intake process for new product partnerships.
• Administrative Productivity: Slackbot handles the administrative tax of the profession, freeing human teams to focus on high-value strategy and counsel. This includes meeting prep and summarization, complex document synthesis, and asset creation, allowing lawyers to start at 80 percent completion rather than a blank page.
• Argument Stress-Testing: Corporate counsel now uses Slackbot to stress-test arguments before they are finalized. By asking the AI to identify logical fallacies or anticipate potential counterarguments, the team is not replacing the lawyer's judgment; it is sharpening it.

How to Build Trust in Agentic Deployments?

As the legal team integrated AI agents into their daily workflows, it became clear they needed a dedicated role at the intersection of legal expertise and agentic risk. Not a blocker, but an enabler with guardrails. This led to the creation of an agent manager role that proactively mitigates legal and reputational risk without becoming a bottleneck for implementation.

The agent manager built and operationalized several key practices:

• Risk and Impact Scoring: A system developed with subject matter experts to prioritize the testing of high-risk queries and high-impact processes.
• User Acceptance Testing Frameworks: Subject matter expert "tiger teams" validate agent accuracy before launch.
• Standardized Guardrails and Disclaimers: Cross-functional teams establish consistent safety boundaries across all deployments.
• Automated Regression Testing: Continuous monitoring of agent performance as models improve, content is updated, and scope increases.
• AI Council Review Checkpoints: A checkpoint in the AI Council Review Process triggers the testing framework whenever a use case is grounded in legal content.

Every agent deployment the agent manager has supported has successfully launched. Having a human sit at the intersection of the legal profession and agentic technology makes agentic deployments successful, with higher trust and adoption within the legal team.

What Does This Pattern Mean for Other Industries?

Legal departments are a fascinating test case for AI transformation precisely because of their historical resistance to change. The professional culture values deliberation, precedent, and risk avoidance. The work involves privileged information, reputational exposure, and significant stakes. The talent pool is, by training, skeptical.

If AI and agentic transformation can take root and thrive in that environment and deliver measurable results, external recognition, and cultural shift, it provides a proof point and a playbook for the entire enterprise. The broader lesson is that agentic systems succeed not when they maximize autonomy, but when they maximize clarity about what agents can and cannot do, who oversees them, and how humans stay in control of high-stakes decisions.

Both the 5G core and the legal department have discovered that the durability of agentic systems depends on open protocols like A2A and MCP, safety policies written as code rather than encoded in prompts, and human oversight positioned as a first-class output rather than an afterthought. These patterns are beginning to define what production-ready agentic AI actually looks like.

" }