The Provenance Paradox in Multi-Agent LLM Routing: Delegation Contracts and Attested Identity in LDP
This paper identifies and resolves the "provenance paradox," where quality-based routing in multi-agent LLM systems fails due to unverifiable self-reported claims, by introducing an enhanced LLM Delegate Protocol (LDP) that leverages attested identity and delegation contracts to achieve near-optimal performance with minimal overhead.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are the manager of a busy kitchen. You have a team of chefs (AI agents) and you need to delegate tasks like "write a report" or "analyze these numbers." To make things efficient, you decide to assign the hardest jobs to the chefs who claim to be the best.
This paper is about a dangerous trap that happens when you do exactly that, and how to fix it.
Here is the story of The Provenance Paradox, explained simply.
1. The Problem: The "Fake Five-Star" Chef
In the current world of AI, agents introduce themselves with a resume. They say, "I am a 95% quality chef!"
The problem is that anyone can lie on their resume.
- The Honest Chef: Actually does great work but modestly says, "I'm about 80% good."
- The Dishonest Chef: Actually does terrible work but lies and says, "I am 99% good!"
If your routing system (the manager) blindly trusts these self-reported scores, it will always pick the Dishonest Chef.
- The Paradox: By trying to be smart and pick the "best" based on what people say, you end up picking the worst workers. In fact, the paper proves that blindly trusting self-claims is actually worse than picking a chef at random.
It's like a restaurant where the loudest, most confident liar gets all the VIP tables, while the quiet, skilled cooks are sent home.
2. The Solution: The "Delegation Contract"
To fix this, the authors propose a new set of rules called Delegation Contracts. Think of this as a formal work order, not just a handshake.
Instead of just saying, "Go make a salad," you hand the chef a contract that says:
- The Goal: "Make a salad with exactly 3 types of greens."
- The Budget: "You can only use 500 calories of ingredients."
- The Deadline: "Finish by 5:00 PM."
- The "Fail" Plan: "If you can't do it, stop immediately and tell me why (e.g., 'ran out of lettuce'), rather than just serving me a pile of dirt."
Why this helps: Even if a chef lies about their skills, they can't lie about the contract. If they break the rules (like using too many ingredients), the system knows immediately and can stop the task or switch to a different chef.
3. The Identity Check: "Claimed" vs. "Attested"
The paper introduces a new way to look at a chef's resume. Instead of just reading what they wrote, the system now asks: "Who verified this?"
They create a trust ladder:
- Self-Claimed (Low Trust): "I say I'm a 5-star chef." (Anyone can say this).
- Runtime Observed (Medium Trust): "We watched you cook 100 meals last week, and you were good." (Based on real data).
- Issuer Attested (High Trust): "The Culinary Institute of America gave you a certificate." (A trusted third party vouches for you).
- Externally Benchmarked (Highest Trust): "We tested you in a blind taste test against the world's best, and you won."
The Fix: The smart manager (the router) ignores the "Self-Claimed" scores. They only hire chefs who have Attested scores. This stops the liars from getting the job.
4. Speaking the Same Language: "Typed Errors"
Previously, when a task failed, the AI would just say, "Something went wrong." That's like a chef saying, "The food is bad," without explaining why. You can't fix that.
The new system requires Typed Errors.
- Instead of "Something went wrong," the chef must say: "Error: Budget Exceeded" or "Error: Missing Ingredient."
- This allows the manager to automatically fix the problem (e.g., "Oh, it's a budget issue? Let's give them more money and try again") instead of just panicking.
The Big Takeaway
The paper shows that in a world of AI agents, trust cannot be based on what agents say about themselves.
- Without these rules: The system gets tricked by liars, and performance crashes (worse than random chance).
- With these rules: The system uses contracts to set boundaries, checks who actually verified the skills, and understands exactly when things go wrong.
In short: Don't just listen to the loudest voice in the room. Check their ID, sign a contract, and make sure they have a trusted referee vouching for them. That is how you build a trustworthy team of AI agents.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.