Builder notes · 13 October 2026 · 6 min read
Support is exact, never by standard
By Anderson and James Warder
- Account abstraction
- Compatibility
- Safe
Why install on the owner's account at all
The easy way to build agent limits is to give every user a new smart account that you control the code of. You get one implementation, one test matrix and no surprises.
We did not do that, for one reason. People already have accounts: a Safe for the treasury, a smart wallet from their app, an EOA they have used for years. Asking them to move funds into a new account in order to make an agent safe adds a migration and a new trust assumption at the exact moment we are trying to remove one. So Gol is one immutable contract, MandateCore, that the owner's existing account authorises to act on its behalf, inside the limits of a mandate.
The agent holds no credential for the account in any of these routes. The account authorised the core. The agent authorised nothing, so it has no path to the funds that skips the check.
Four routes into an account
| Route ("lane") | How the owner installs it | Account families supported today |
|---|---|---|
| ERC-7579 executor | installModule as an executor | Nexus 1.0.0, Kernel v0.3.1, an EOA delegated to Nexus 1.3.3 under EIP-7702 |
| Safe module | enableModule, through the Safe's own threshold | Safe 1.4.1 (L2 singleton) |
| ERC-6900 validation | installValidation | Alchemy Modular Account v2 |
| ERC-20 allowance | approve at the token | Plain EOAs. Designed, not in the current profile |
Where "same standard" stopped meaning "same behaviour"
Modular-account standards exist so that one module can run on many accounts, and mostly they deliver. But when we installed the same core through the same lane on two accounts that both conform to the standard, they disagreed on what the installation data had to contain, and on how an owner signature is formed and checked. Both behaviours are defensible readings of the standard. They are not the same behaviour.
For an ordinary module, that is a compatibility bug you fix when someone reports it. For a module whose whole job is to decide whether money may move, it is a security boundary. If we assumed support from a standard's name, we would be claiming safety on accounts we had never actually tested.
So we wrote the rule down:
We publish that list as a compatibility profile: an immutable record of the exact network, core, accounts, actions, rule kinds, assets and providers supported in one release. A mandate commits to the profile it was approved under, so a later release can never quietly reinterpret what an owner agreed to. Research notes, docs and adapter code do not count as support. Only a profile does.
Three things that are harder than they look
1. Upgrades you cannot see. If the account's implementation, delegation, validators, hooks or executors leave the reviewed configuration, agent execution stops and refuses with account_integration_invalid until the owner restores it. That part is on-chain. But on Kernel v0.3.1 and Alchemy Modular Account v2, the core cannot detect a same-identity implementation upgrade from inside the chain. For those accounts our platform checks the implementation slot before every approval and every execution. That check is off-chain, and we say so in the docs. It is the kind of caveat that is easy to leave out, and readers notice when it is missing.
2. Removals that do not tell you. Removing Gol from an account should disconnect execution. Retiring it permanently should also make sure no earlier signature can ever be replayed. We do that with an account-wide invalidation that advances an epoch. On Safe and on Alchemy, the native removal path gives the module no callback. So the SDK prepares invalidation and removal as one atomic owner transaction, and we test that bundle like any other rule. It is not left as a convention.
3. Allowances that outlive the mandate. On the plain-EOA lane, the owner approves the core to pull a token. If the owner later revokes the mandate, the token allowance stays until they clear it at the token. The core refuses every pull outside an active mandate and cannot be upgraded. The allowance is still there, though, and we will not ship that lane until the approval screen and the docs state it plainly.
The cost of being immutable
MandateCore cannot be upgraded. That is the point: no admin key can widen what an owner signed. It also means:
- The contract is roughly 1.1 to 1.5 KB under the EVM's 24 KB contract size limit (EIP-170). A fix that needs more room ships as a new core, and every owner has to install it and sign again, because the verifying contract is part of what they signed. Every change runs a size check, and the migration cost is written down before the change is accepted.
- New rule kinds and actions enter through append-only registries under new numeric tags. A mandate commits to the exact tags it was approved with, so adding a registration can never widen an existing grant.
- Known gap: the registries cannot yet deprecate an entry. A mistaken registration is permanent on that core. We list that in our open issues rather than hope nobody asks.
What we would tell another team
- Test the account, not the standard. Keep a matrix of exact implementations and versions, and treat each new row as a release decision.
- Make your support claim a document that cannot change. If an owner's approval can be reinterpreted by your next release, you have built a server with extra steps.
- Write down every check that runs off-chain. Off-chain checks are fine. Hiding them is not.
Where we stand
Live on Base Sepolia under compatibility profile 4, with 28 contracts verified on Sourcify. Unaudited. One action today: USDC transfers to recipients the owner listed. Mainnet waits for an audit.
Disclosure: Morca Labs builds Gol. This post describes Gol's own design decisions.