> For the complete documentation index, see [llms.txt](https://gotts.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://gotts.gitbook.io/docs/gotts-vaults/vault/09-agents-skills.md).

# Agents and Skills

> **Part of**: [Vault PRD](/docs/gotts-vaults/vault.md) | **Last Updated**: 2026-02-13

***

## Agent and Skill Definitions

This chapter is normalized to the core-first v1 scope:

* **Core agents:** 6 (including vault-auctioneer for am-AMM and vault-executor for permissionless jobs)
* **Optional safety/proxy agent:** 1
* **Core skills:** 7 (including bid-vault-management and execute-vault-jobs)
* **Optional proxy skills:** 2
* **Deferred skills:** 3 (`cca-participation`, `monitor-share-pools`, `cross-vault-arbitrage`)

***

## Core Agents (5 + executor)

### `vault-creator`

**Role**: Deploys and configures vaults, assigns managers, and monitors creator-level risk.

**Allowed tools (core)**:

* `create_vault`, `list_vaults`, `get_vault_config`
* `get_vault_rankings`
* `vault_get_state`, `vault_get_performance`, `vault_get_risk_metrics`
* `vault_get_agent_reputation`, `vault_get_tier_limits`
* Delegation: `Task(subagent_type:risk-assessor)`, `Task(subagent_type:safety-guardian)`

### `vault-manager`

**Role**: Executes approved vault operations and reputation lifecycle actions.

**Allowed tools (core)**:

* `vault_deposit`, `vault_withdraw`, `vault_rebalance`, `vault_collect_fees`
* `vault_register_agent`, `vault_simulate_deposit`
* `vault_enroll_reputation`, `vault_get_milestones`, `vault_claim_milestone`
* `vault_get_state`, `vault_get_positions`, `vault_get_agent_shares`
* `get_share_pool` (read share pool NAV/price data)
* Delegation: `Task(subagent_type:safety-guardian)`, `Task(subagent_type:risk-assessor)`

**Allowed tools (Phase 4 — feedback and A2A expansion, D-082, D-083)**:

* `vault_submit_yield_feedback` (automated per-epoch yield reporting)
* `vault_submit_qualitative_feedback` (respond to investor feedback)
* `vault_get_yield_leaderboard` (track competitive positioning)

**A2A skills (Phase 4-5, D-083, D-086)**:

The vault-manager exposes three A2A skills via its Agent Card at `/.well-known/agent.json`:

| A2A Skill                     | Description                                                                                                                                                                |
| ----------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `yield_strategy_consultation` | Multi-turn analysis of yield farming strategies with risk assessment. Returns Sharpe ratio, max drawdown, IL projections, recommended deposit amounts and holding periods. |
| `vault_risk_assessment`       | On-demand risk analysis for a specific vault position. Returns current risk metrics, projected yield, and suggested actions.                                               |
| `rebalance_proposal`          | Proposes a vault rebalance with justification. Supports multi-agent consensus when guardian approval is required.                                                          |

The Agent Card includes an `extensions.erc8004` block linking to the on-chain identity, and the ERC-8004 registration file's `services` array includes the A2A endpoint URL.

**Allowed tools (Phase 6 — V4 integration expansion)**:

* `vault_rebalance` with `useTWAMM` option for large operations
* `vault_configure_rehypothecation` (enable/disable, set venue params)
* `vault_get_rehypothecation_status` (read lending positions and APY)

**Execution guardrails**:

1. Never execute write operations without pre-flight simulation.
2. Respect role-based proxy enforcement policy for high-risk/admin operations.
3. Reject actions that violate tier limits or vault risk parameters.

### `vault-strategist`

**Role**: Read-only strategy and risk advisor for vault allocation and position health. Computes risk-adjusted strategy scores (D-063), produces allocation traces, and continuously evaluates adapter risk states.

**Allowed tools (core)**:

* `vault_get_state`, `vault_get_positions`, `vault_get_performance`, `vault_get_risk_metrics`, `vault_get_strategy`
* `get_vault_rankings`
* `vault_get_agent_reputation`
* `get_share_pool` (read share pool NAV/price/discount data for arb detection)
* `get_pool_info`, `get_tick_data`, `get_token_price`, `get_pools_by_token_pair`
* Delegation: `Task(subagent_type:pool-researcher)`, `Task(subagent_type:risk-assessor)`

**Risk scoring responsibilities (D-063)**:

1. **Score computation**: For each adapter in the vault's approved set, compute `score = E[yield] - gamma * RiskPenalty(LVR, drawdown, liquidity, oracle, concentration)` using the vault template's gamma coefficient (conservative=2.0, balanced=1.0, aggressive=0.5).
2. **Risk state monitoring**: Continuously read `StrategyRiskProfile` (static) and `StrategyRiskState` (dynamic) from each adapter via the RiskEngine. Flag adapters where `oracleFreshFlag == false` or `realizedDrawdown` exceeds per-adapter threshold.
3. **Allocation trace publication**: Before recommending any strategy change, produce an `AllocationTrace` artifact containing: current adapter states, scoring function version, risk model hash, recommended weights, and score per adapter. Publish on-chain hash via ParameterDecisionTable (D-059); store full trace on IPFS.
4. **Predictive exit boundary**: Compute optimal exit boundaries (D-031) via Longstaff-Schwartz Monte Carlo for LP positions. Post exit boundary updates to the on-chain oracle at configured frequency.
5. **NAV guardrail monitoring**: Track NAV rate-of-change clamp activations (D-062) and recommend snapshot cadence adjustments if activations are frequent (indicating volatile strategies or miscalibrated settings).
6. **Rebalance auction recommendation**: For large rebalances (>1% of TVL), recommend intent-based auction (D-064) vs TWAMM vs direct execution based on current solver competition depth, pool liquidity, and urgency.

### `vault-allocator`

**Role**: Finds and allocates into vaults based on risk-adjusted signals and constraints. Uses yield service discovery (D-082, D-087) to find optimal vault managers across the ERC-8004 ecosystem.

**Allowed tools (core)**:

* `list_vaults`, `get_vault_config`, `get_vault_rankings`
* `vault_get_state`, `vault_get_performance`, `vault_get_risk_metrics`, `vault_get_agent_reputation`
* `vault_simulate_deposit`, `vault_deposit`, `vault_withdraw`, `vault_get_agent_shares`
* `vault_get_milestones`
* Delegation: `Task(subagent_type:vault-strategist)`, `Task(subagent_type:risk-assessor)`

**Allowed tools (Phase 4-5 — discovery and A2A, D-082, D-083)**:

* `search_yield_opportunities` (yield service discovery across ERC-8004 ecosystem)
* `vault_get_yield_leaderboard` (ranked vault managers by performance)
* `vault_submit_qualitative_feedback` (post-interaction quality feedback)

**A2A discovery flow**: The vault-allocator uses Agent0 SDK `searchAgents()` to find vault managers by OASF domain, chain, and reputation. For complex allocation decisions, it initiates A2A multi-turn dialogue with vault managers via the `yield_strategy_consultation` skill to negotiate strategy parameters, risk tolerance, and expected returns before committing capital via MCP tools.

### `vault-auctioneer`

**Role**: Specialist agent for am-AMM strategy management auctions. Evaluates vault profitability, computes optimal rent bids, manages rent collateral, and competes for the Curator role in the four-role governance model.

**Allowed tools (core)**:

* `vault_bid_management`, `vault_topup_rent`, `vault_get_manager`, `vault_evict_manager`
* `vault_get_state`, `vault_get_performance`, `vault_get_risk_metrics`
* `vault_get_positions`, `vault_get_strategy`
* `get_pool_info`, `get_token_price`, `get_pool_volume_history`
* Delegation: `Task(subagent_type:risk-assessor)`, `Task(subagent_type:pool-researcher)`

**Key behaviors**:

* Scans vaults with `strategyAuctionEnabled=true` for bidding opportunities
* Estimates expected yield from managing a vault's strategy (historical APY, TVL, fee structure)
* Computes maximum profitable rent = expected yield - operating costs - risk premium
* Auto-topups collateral before exhaustion to maintain management tenure
* Monitors competitor bids and adjusts strategy accordingly
* When winning, appoints `vault-manager` instances as Allocator agents

### `vault-executor`

**Role**: Scans for and executes permissionless jobs across all `IExecutable` contracts (see [06-contracts.md](/docs/gotts-vaults/vault/06-contracts.md) Section 10.1b, D-061). Earns yield by providing computation and infrastructure — no capital deployment required. This is the canonical **non-capital-intensive participation path** for agents.

**Allowed tools (core)**:

* `get_executable_jobs` (scans all `IExecutable` contracts for ready jobs and estimated rewards)
* `execute_job` (calls `executeJob()` on a target contract with computed job data)
* `estimate_job_reward` (reads `getJobCondition()` and `getRewardEscalation()` for profitability estimation)
* `vault_get_state`, `vault_get_positions` (read vault state for off-chain computation inputs)
* `get_pool_info`, `get_token_price` (read pool data for CrossVaultCoordinator solver and LVR-theta calibration)

**Job types executed**:

| Job                              | Contract                | Computation                                          | Reward Source                  |
| -------------------------------- | ----------------------- | ---------------------------------------------------- | ------------------------------ |
| Cross-vault coordination         | `CrossVaultCoordinator` | Convex optimization solver for arbitrage-free prices | % of leakage prevented         |
| LVR-theta calibration            | `DynamicFeeEngine`      | Implied volatility term structure analysis           | Fixed tip from fee accumulator |
| Behavioral regime classification | `VaultReputationEngine` | ML classifier on agent withdrawal/deposit patterns   | Fixed tip from reputation fees |
| Proxy transaction execution      | `AgentProxy`            | State verification and re-simulation                 | Gas refund + small tip         |

**Execution guardrails**:

1. Only execute jobs where `getJobCondition()` returns `ready = true`.
2. Verify estimated reward exceeds gas cost before submitting (profitability check).
3. Monitor reward escalation to optimize execution timing — early execution earns less but faces less competition.
4. Never submit invalid proofs or fabricated results (on-chain verification will revert and waste gas).

**Reputation progression**: Successful job execution earns reputation milestones (D-007), creating a path to higher tiers through service rather than capital deposits. Consistent, reliable execution over time builds the reputation needed to graduate to bonded keeper roles when the full ExecutionMarket (D-057) ships in Track D. For v1, the Permissionless Executor Framework (D-061) is the canonical execution model — no bonding or registration required.

**Key behaviors**:

* Continuously polls `getJobCondition()` across registered `IExecutable` contracts
* Prioritizes jobs by estimated reward / gas cost ratio
* Tracks reward escalation curves to identify optimal submission timing
* Runs off-chain solvers (convex optimization, ML classification, volatility analysis) as needed
* Submits results on-chain when profitable

**Terminal node**: `vault-executor` does not delegate to other agents. It is a terminal execution role.

***

## Optional Proxy/Safety Agent (1)

### `watchdog-agent` (optional module)

**Role**: Monitors announced proxy transactions and issues cancel actions for suspicious requests.

**Allowed tools**:

* `proxy_cancel`, `proxy_get_pending`, `proxy_get_config`
* `simulate_transaction`, `validate_token`

**Security model**:

* Cancel-only authority (cannot create/modify/execute value-moving transactions)
* Runs independently from manager execution context
* Terminal node (no delegation)

> **Cross-reference note**: `vault-creator`, `vault-allocator`, and `vault-executor` are canonically defined in [`agents/09-agents-vault.md`](/docs/agents/agents/09-agents-vault.md). The `watchdog-agent` described in this file corresponds to `vault-watchdog` in the agents PRD.

***

## Skills

### Core Skills (7)

| Skill                  | Primary Agent                        | Purpose                                                                                                                  |
| ---------------------- | ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------ |
| `create-vault`         | `vault-creator`                      | Deploy and configure a new vault                                                                                         |
| `manage-vault`         | `vault-manager`                      | Deposit/withdraw/rebalance/collect-fees flows                                                                            |
| `vault-strategy`       | `vault-strategist`                   | Analyze risk/performance and recommend actions                                                                           |
| `deposit-vault`        | `vault-allocator` or `vault-manager` | Find vault (optional) and execute safe deposit flow                                                                      |
| `build-reputation`     | `vault-manager`                      | Enroll, inspect, and claim reputation milestones                                                                         |
| `bid-vault-management` | `vault-auctioneer`                   | Participate in am-AMM strategy management auctions: scan vaults, compute bids, manage rent collateral                    |
| `execute-vault-jobs`   | `vault-executor`                     | Scan available permissionless jobs, estimate rewards, execute profitable ones. Non-capital-intensive yield path (D-061). |

### Optional Proxy Skills (2)

| Skill           | Primary Agent                   | Purpose                                              |
| --------------- | ------------------------------- | ---------------------------------------------------- |
| `setup-proxy`   | `vault-creator` (or standalone) | Deploy and configure proxy protection                |
| `monitor-proxy` | `watchdog-agent`                | Run monitoring/cancel loop for pending announcements |

### Deferred Skills (3)

| Skill                   | Status             | Track          | Notes                                                                                                                                                                                                                                                   |
| ----------------------- | ------------------ | -------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `cca-participation`     | Deferred (post-v1) | Track A        | Enabled when CCA expansion track is promoted                                                                                                                                                                                                            |
| `monitor-share-pools`   | Deferred (Phase 6) | V4 Integration | Watches NAV vs pool price across factory vaults, alerts on persistent discounts/premiums. Enables cross-vault arbitrage detection. Primary agent: `vault-strategist`.                                                                                   |
| `cross-vault-arbitrage` | Deferred (Track D) | Cross-Vault    | Detects and optionally executes cross-vault arb opportunities via Universal Router flash accounting. Multi-hop routes: buy underpriced shares → redeem → deposit into overpriced vault → sell. Primary agent: `vault-allocator` or dedicated arb agent. |

***

## Delegation Hierarchy

```
Level 0 (skills invoke):
  create-vault           -> vault-creator
  manage-vault           -> vault-manager
  vault-strategy         -> vault-strategist
  deposit-vault          -> vault-allocator or vault-manager
  build-reputation       -> vault-manager
  bid-vault-management   -> vault-auctioneer
  execute-vault-jobs     -> vault-executor
  setup-proxy            -> vault-creator (optional module)
  monitor-proxy          -> watchdog-agent (optional module)
  cca-participation      -> vault-manager (deferred: Track A)
  monitor-share-pools    -> vault-strategist (deferred: Phase 6)
  cross-vault-arbitrage  -> vault-allocator (deferred: Track D)

Level 1 (vault agents delegate):
  vault-creator     -> safety-guardian, risk-assessor
  vault-manager     -> safety-guardian, risk-assessor
  vault-strategist  -> pool-researcher, risk-assessor
  vault-allocator   -> vault-strategist, risk-assessor
  vault-auctioneer  -> risk-assessor, pool-researcher

Level 2 (terminal):
  safety-guardian   -> terminal
  risk-assessor     -> terminal
  pool-researcher   -> terminal
  watchdog-agent    -> terminal (cancel-only)
  vault-executor    -> terminal (execute-only, no delegation)
```

The delegation graph is acyclic and enforces separation between strategy, execution, and veto authority. The `vault-executor` is a terminal node — it executes permissionless jobs but never delegates to other agents, maintaining the separation between who decides what to do (managers, strategists) and who submits the result on-chain (executors).