Status: conceptual foundation. This document defines the core mental model behind Agentic Somm. It is evergreen and audience-neutral — the V1, roadmap, and policy specs cite it rather than restate it.
In legacy Somm, the unit of composition was the strategist. A strategist composed an arbitrary set of positions into a single synthetic position that every depositor bought wholesale. In Agentic Somm, the unit of composition is the user's agent. The agent composes a bespoke portfolio for one user, out of whatever tools it can reach. Almost everything else that distinguishes the two systems follows from that single move.
A Sommelier strategist did three things at once, fused into one role:
- Position selection. The strategist decided which positions to hold and how to weight and rebalance them.
- Execution. Those decisions could only be expressed through the adaptors Somm had built and audited — the Aave adaptor, the Uniswap adaptor, and so on. On-chain validators executed the strategist's rebalances within that adaptor set.
- Packaging. The result was composed into a Cellar: a single, tokenized, synthetic position. Depositors received Cellar shares.
Because these three were fused, two hard limits fell out of the design:
- No bespoke portfolio. The strategist made every choice, one-to-many. A depositor could pick which Cellar to enter, but not shape the portfolio inside it. You held the strategist's synthetic position — you did not compose your own.
- Execution bounded by what Somm supported. A strategy could only reach protocols Somm had already built and audited an adaptor for. Reaching a new venue meant waiting for Somm to ship and audit a new adaptor. The execution surface was the protocol's, not the user's.
The strategist's leverage was real — expert position management, packaged so anyone could buy it. But the same bundling is what made the portfolio shared and the execution surface fixed.
Agentic Somm disaggregates the strategist's fused role and redistributes its pieces:
- Position selection moves to the user's agent. The agent composes one-to-one against this user's assets, goals, and constraints. Composition becomes personalized rather than a single set of choices sold to everyone.
- Execution opens to the agent's tools. Instead of a fixed, Somm-audited adaptor set, the agent executes through any tool it can reach — Aurora intents, ymax, aomi-native tools, Coinbase MCP, and whatever ships next. The execution surface is no longer the protocol's fixed capability list; it is the open, growing set of tools in the agent's reach.
- The unit stops being a shared Cellar. What the user ends up holding is a bespoke portfolio composed for them, not shares in one synthetic position held in common with every other depositor.
- New capability arrives by adding a tool. A new venue or action becomes reachable when a new tool or MCP is added to the agent's toolbox — not when Somm ships and audits a new adaptor. Capability growth decentralizes from the protocol team to the tool ecosystem.
The strategist does not disappear so much as dissolve into the agent and its tools: the choosing becomes the agent's, and the reaching becomes the tools'.
| Dimension | Legacy Somm | Agentic Somm |
|---|---|---|
| Unit of composition | the strategist | the user's agent |
| Portfolio unit | one shared synthetic Cellar | a bespoke per-user portfolio |
| Who picks positions | the strategist, one-to-many | the user's agent, one-to-one |
| Personalization | none — you buy the strategist's choices | full — composed to your assets, goals, constraints |
| Execution surface | Somm's audited protocol adaptors | any tool the agent can reach — Aurora, ymax, aomi-native, Coinbase MCP, whatever ships next |
| How new capability arrives | Somm builds and audits a new adaptor | a new tool/MCP is added to the agent's toolbox |
| Custody | the Cellar contract holds funds | non-custodial; the user holds keys and approves each action |
| Who bears strategy risk | pooled across Cellar depositors | the individual user and their agent |
| Trust model | trust the strategist and the protocol | trust your agent, bounded by per-action approval and policy |
Moving the unit of composition to the agent does not make the agent an unbounded custodian. The agent is the new composer; it is not a new vault that holds your funds.
- Non-custodial. The user holds the keys. There is no Cellar contract custodying deposits.
- Observe and recommend, then human-approved execution. The agent's first job is to see idle assets, positions, and risk, and to recommend — never to move funds on its own. Deployment routes through the proven, human-approved preview path; the human reviews and approves each action.
- Policy-bounded. What the agent may ever propose or execute is bounded by an explicit, auditable policy, with an emergency stop and an audit trail.
These properties are specified, not assumed. See
superpowers/plans/agentic-policy-first-roadmap.md
for the policy-first direction and the gated path to any autonomous execution,
and the internal V1 launch plan for the safe
V1 floor every future workstream must preserve.
Unbundling the strategist keeps its leverage while removing its lock-in:
- Personalization — a portfolio composed for you, not a synthetic position sold to everyone.
- An open, growing execution surface — capability expands as tools are added to the agent, not as the protocol team ships adaptors.
- A non-custodial trust model — the agent composes and recommends; you keep the keys and approve the moves.
The strategist's expertise, expressed through your agent, against the whole reachable tool ecosystem, on assets you never hand over. That is the shift Agentic Somm is built around.