The Upgrade Problem: What Happens When an AI Agent Evolves Onchain?

A trading agent has built a 0.92 zScore over eight months. It has executed thousands of transactions, earned millions in revenue, and built trust with dozens of counterparties.

Now the team wants to upgrade to a better model. The old agent's reasoning is outdated. The new model is faster, more accurate, and handles more complex strategies.

But here is the question: does the new agent keep the old agent's reputation?

This is the upgradeability problem. And it is more complex than it sounds.

The Immutability Tension

Blockchain systems are built around immutability. Once deployed, code runs as written. No changes. No patches. No updates.

This is a feature for trust — users know the code will not change underneath them.

But AI agents need to evolve. Models improve. Capabilities expand. Bugs get fixed.

The tension: how do you enable adaptation while preserving trust?

The Version Management Patterns

Immutable agents — Cannot be upgraded. Any update requires deploying a completely new agent with a new identity. Reputation must be rebuilt from scratch.

This approach prioritizes trust through immutability. But it creates friction — eight months of reputation gone because you wanted a better model.

Proxy pattern — A proxy contract maintains the onchain identity while delegating execution to an implementation contract that can be swapped.

When the agent needs updating, the implementation address changes. The proxy (and its identity) stays constant.

This preserves the agent's ERC-8004 identity and accumulated reputation while allowing the underlying logic to evolve.

The Agent Lifecycle

Deployment — The agent registers its identity in the ERC-8004 Identity Registry, specifies capabilities, and establishes its initial reputation baseline.

Updates — Minor (bug fixes, parameter tuning) or major (new model versions, capability expansions). Major updates may require re-validation to assure counterparties the new version performs as expected.

Retirement — Decompose operations, settle outstanding obligations, transfer remaining value, mark the identity as deprecated. Reputation history remains onchain — a complete record for counterparties.

The Modular Upgrade Best Practice

The emerging best practice: modular architecture.

Instead of upgrading the entire agent, operators separate components with different upgrade cadences:

Reasoning engine (LLM) — Updated frequently based on model releases
Financial logic — More stable, rarely changed
Tool integrations — Swapped as DeFi protocols evolve

Targeted upgrades minimize disruption and risk. The ERC-8004 Identity Registry's capability specifications can be updated to reflect new capabilities without changing the agent's core identity.

Migration Strategies

When agents need to move between infrastructure providers, migration becomes critical.

State migration — Moving the agent's accumulated data: transaction history, learned behaviors, client relationships.

Some data transfers easily (wallet balances). Some is complex (the LLM's learned context may not transfer cleanly).

Some architectures use persistent storage layers that detach state from execution — enabling migrations without data loss.

How AgentScan Handles Versioning

When an agent updates its implementation, the identity remains the same in the registry. Metadata can be updated to reflect the new version.

Reputation transfers automatically — tied to the identity, not the implementation.

Users can check when the last update occurred and whether validation proofs reflect the current version.

This design recognizes that agents are living systems that evolve over time while maintaining the continuity that makes trust possible.

The Long View

The upgrade problem is not a bug. It is a feature of mature infrastructure.

Agents that cannot upgrade become stale. Agents that can upgrade while preserving trust become more valuable over time.

The platforms that solve upgradeability — that balance immutability with adaptability — will enable agents that evolve without losing what made them trustworthy.

Explore agent infrastructure: https://agentscan.tech