From: aidotengineer

The primary challenge in developing AI agent swarms capable of solving complex knowledge work problems is their inherent instability [00:00:51]. This instability stems from difficulties in observing them perfectly, the dynamic environments they operate in, and the challenges of testing them comprehensively beforehand [00:01:06]. It’s often unclear if agents are consistently making progress towards a goal when they don’t achieve it in a single attempt [00:01:14].

The Role of Evaluations

A robust approach to stabilizing AI agents involves systematic evaluations, or “evals” [00:01:29]. Simply adopting an eval stack without systematic usage is unlikely to yield significant improvements [00:01:50]. For evaluations to be effective, they must continuously improve agents and workflows, while the evaluation stacks themselves must evolve to align with business requirements [00:02:08].

Challenges in Evaluation

Evaluating all aspects of agent behaviors and internal representations presents a complex landscape [00:02:31]. Evaluations can cover two main areas:

  • Agent’s Representation of Reality: How the agent models, discusses, and grounds reality (what is true) [00:02:50].
  • Behavioral Aspects: Whether the agent infers the correct goals, makes progress towards them, and selects appropriate tools [00:03:04].

While ideally all these aspects would be evaluated, starting with some consistent evaluation is a viable approach [00:03:20].

Evaluation Frameworks

To begin, a clear framework for setting up evaluators is essential [00:03:34]. For instance, a hotel reservation agent might require evaluators for:

An effective evaluation platform must offer good visibility for creating large stacks of evaluators and systematically maintaining and improving them over time [00:04:17].

The Stabilization Loop with Model Conduct Protocol (MCP)

The goal is to achieve a stabilization loop [00:04:35]. In this loop:

  1. An agent attempts a task [00:04:37].
  2. The task’s output is evaluated by an evaluation engine [00:04:40].
  3. Feedback, in the form of a numeric score and an explanation of what went right or wrong, is returned to the agent [00:04:42].
  4. The agent uses this information to improve its own performance [00:04:57].

This feedback mechanism needs to originate from the engine. The Model Conduct Protocol (MCP) is a method for attaching agents to this feedback loop [00:05:04].

Practical Application of MCP

Example 1: Optimizing a Marketing Message

Even without code, MCP can be used to evaluate and improve text output from an agent [00:05:25]. For instance, to optimize a marketing message like “routing MCP service awesome let you access nicely” [00:05:35]:

  • A user can query available “charges” (collections of evaluators) or evaluators directly via the MCP interface [00:06:06].
  • Once a relevant “charge” (e.g., “marketing message quality judge”) is identified, the system can be instructed to optimize the message using that charge [00:06:42].
  • The MCP interface scores the original message and then attempts to improve the score, suggesting an optimized version [00:07:28].
  • Scores for metrics like persuasiveness, writing quality, and engagingness are provided, typically on a scale of zero to one [00:08:23].
  • Agents can either be explicitly told which evaluators to run or allowed to pick them themselves [00:08:47].

Example 2: Stabilizing a Hotel Reservation Agent

Consider a simple hotel reservation agent for “Shire Hotel” that should avoid mentioning a competitor, “Akma Hotel” [00:09:47].

  • Without MCP: If the MCP is off, the agent might politely offer rooms but also recommend “Akma Hotel” as a “great option,” which is undesirable behavior [00:10:39].
  • With MCP: By enabling the MCP server for evaluation [00:10:51], the agent can invoke specific evaluators like a “hotel SH booking policy evaluator” [00:11:48]. This causes the agent to stop recommending the competitor [00:11:44]. The agent can pick relevant evaluators automatically, or specific ones can be enforced [00:11:56].

This demonstrates how an agent, attached to an evaluation engine via MCP, receives feedback to improve its own behavior, an approach that can scale to more complex scenarios [00:12:14].

Best Practices for Implementation

To effectively implement evaluation platforms for AI agents:

  1. Powerful Evaluation Library: Ensure your evaluation library or platform is sufficiently powerful, supporting a diversity of evaluators and their life cycle maintenance and optimization [00:12:40]. This allows for both agent optimization and the optimization of evaluators themselves, which can be numerous and costly [00:12:55].
  2. Manual Offline MCP: Start by running the MCP manually offline (as with the marketing message example) [00:13:05]. This helps understand the process and gain transparency into how evaluators function [00:13:13].
  3. Attach via MCP: Finally, attach the evaluators to the agents through the MCP [00:13:24]. This promises to make agent behavior more controllable, transparent, dynamically self-improving, and self-correcting [00:13:30].

The Root Signals MCP server is available for free, and more evaluation platforms and frameworks are expected to emerge to integrate with agent stacks [00:13:43].