The skills required to manage your workforce of AI agents

By Gaurav Suman, Director of AI and Industry Solutions, Solace

The software industry has long understood that reliable systems require a structured, repeatable way of building them. Without a disciplined approach to how software is planned, built, tested, and maintained, quality erodes, and systems become impossible to scale – yielding software that’s fast at first but brittle forever. This truth gave birth to the software development lifecycle (SDLC), a multistep framework that codified methodology as a core ingredient of reliable software delivery. 

AI agents are also software, but not software “as we know it.” Built on probabilistic language models, large or small (large language model/small language model or LLM/SLM), there are fundamental differences in how agents are built, evaluated, and governed. Introducing them into an organisation needs an agent development lifecycle (ADLC) applying SDLC’s philosophy of structured, repeatable process to the development and management of AI agents. 

Source: Solace. Suman.

With help from HR practices 

In Singapore, Singtel is redesigning its organisational structure where managers oversee mixed fleets of humans and AI agents. Meanwhile, GovTech is developing a registry to track the owners and activities of AI agents used by 150,000 public officers. These developments point to a simple idea: since AI agents perform roles within an enterprise, the ADLC should also draw from the human employee lifecycle. 

When an organisation hires a new team member, there’s a deliberate process: their role is defined; access is granted to systems they need; they’re trained and integrated into a team. Once empowered, a manager oversees and monitors their performance over time. As autonomous actors operating within a system of governance, AI agents deserve and require that same treatment; they need real-time contextual data to do their jobs! 

Don’t skip the AI “hiring process” 

Picture an organisation where every employee is a generalist, with no hierarchy or structure, communicating only via synchronous one-to-one phone calls – no emails, Slack, or shared systems. That organisation can’t scale. Information doesn’t flow. People drown in data, with no real understanding of what to do. There are no accountability structures, access controls, or coordination mechanisms beyond individual conversations.

You wouldn’t drop a new human hire into this chaos. Effective organisations instead employ specialised people, organised hierarchically, who coordinate asynchronously. Onboarding AI agents requires that same approach.

Most early agentic AI deployments work as complex monolithic agents tasked with doing everything, accessing all data, or maintaining large context with sub-agents. The result? A system that’s brittle, expensive, inconsistent, and unable to handle real enterprise complexity.

This is where an agent mesh acts as a development and runtime platform, helping build AI agents for a real-time enterprise, specialised for particular functions, organised hierarchically, and orchestrated by a development and runtime platform that delegates tasks to appropriate agents. Communications between agents are asynchronous and event-driven, while role-based access controls give agents exactly the permissions they need and nothing more.

So, let’s look at how agents can be “onboarded” through an ADLC much like a new employee joining the organisation: 

Hiring: Define the role, responsibilities, expectations, guardrails 

Before a human employee starts, you write the job description, defining their role, responsibilities, expected behaviours, and boundaries. An agent mesh’s builder uses an internal AI agent to guide the setup of an agent’s purpose, scope, and configuration. You set the instructions and system prompts shaping the agent’s persona and configure guardrails – hard constraints that prevent it from going off-script or taking unsafe actions.  

Onboarding: Give access to systems and tools 

A new employee’s first weeks are spent getting access to tools, systems, and data. Agent onboarding is the same. An agent mesh should include pre-built integrations to enterprise databases, data warehouses, data lakes, APIs, and Model Context Protocol (MCP) servers. Role-based access controls (RBAC) enforce least-privilege principles so agents only see what they need, while defined skills help agents use their tools well. 

Coaching: Internal training to ensure competence 

After getting access to systems, human employees undergo training to turn general ability into job-specific skills. The same logic applies here. An agent mesh can provide an Eval function* that uses AI to suggest tests for agents, lets you add more and runs them against your agents. This empowers the organisation to validate agent competence against defined success criteria, supporting both initial and regression testing as you make changes. 

Supervision: Trust, but verify 

Even capable employees get close oversight when new to a role. Supervision isn’t micromanagement; it’s the safety net that ensures quality and catches errors before they compound. An agent mesh with human-in-the-loop architecture can route specific agent actions or decisions to human reviewers before execution. This matters as LLMs are non-deterministic** and make mistakes, so when the impact of agents being wrong is too risky, humans can validate their actions, responses or conclusions.  

Teamwork: Where the real value emerges 

The most transformative phase of the employee lifecycle is when individuals become part of high-performing teams, where collective capability exceeds the sum of its parts. An agent mesh can support repeatable workflows, such as steps in approving a loan or providing an insurance quote. Dynamic orchestration is where the mesh comes alive: orchestrator agents route work to the right specialists, in sequence or in parallel, adapting in real time when the path forward requires reasoning.

The result is a multi-agent mesh topology – hierarchical agent organisations coordinating specialist agents across various functions, mirroring the structure of an effective human organisation. 

Improvement: Deployment is not the finish line – it’s the beginning

Effective organisations don’t deploy employees and forget about them. They monitor performance and provide feedback for continuous improvement.

An agent mesh’s visualiser*** gives a real-time graphical interface for tracing agent interactions, tool and LLM calls, and decision pathways. Ongoing evaluations detect performance drift**** or emerging failure modes via online evals in the background, monitoring production execution. Further, OpenTelemetry instrumentation surfaces performance trends, giving you data to make informed decisions about when and how to tune your agents.

Managing agents like employees

Moving AI agents from experimental prototypes to mission-critical enterprise assets requires a fundamental shift in how we think about agent delivery. SDLC provides discipline for deterministic code, but autonomous, probabilistic agents require more: clear roles, secure system access, rigorous testing, human supervision, and continuous performance management.

By combining the structural discipline of the ADLC with a real-time and event-driven platform, enterprises can confidently deploy, orchestrate, and govern multi-agent systems at scale.

*An eval function runs code and produces the result from running the code.

**When something is deterministic, the outcome is always the same for a given situation. When it is non-deterministic, the same situation can lead to different outcomes. 

*** A visualiser provides visibility into live connections or work.

****Performance drift refers to deviation from expected outcomes.