Agent custody and authorization framework · 2026-05 reference architecture

Confidence: Likely Updated 2026-05-25 Review by 2026-11-25 Sources 14 Machine-translated Original (JA)
#agent-economy#custody#authorization#revocation#multi-sig#threshold
On this page

TL;DR

Agent custody and authorization in 2026 has converged into a three-axis design space: (1) custody model — full self-custody (user holds key, agent has no independent control), hybrid (Lit PKP threshold or session-key with policy code), or fully custodial (Privy / Magic / Coinbase CDP / Skyfire); (2) authorization granularity — single-transaction (agent must request user approval per action), time-bounded (scope valid until expiry), amount-bounded (spend cap with rolling window), spend-rule-scoped (allowed merchants / categories / patterns); (3) revocation mechanism — on-chain instant (burn permission NFT, revoke ERC-7715 grant), off-chain wallet-mediated (Privy admin revoke), legal / contractual (Skyfire issuer ledger entry). The traditional brokerage limited power of attorney (LPOA) provides the doctrinal analog — for decades, brokerages have given investment advisers discretionary authority over client accounts with scope (asset classes), amount (size limits), revocation (written notice), and audit trail (statements). The 2026 agent custody architecture is digitally-native LPOA plus on-chain revocation primitives. Multi-signer / threshold control adds a second-pair-of-eyes pattern: a transaction requires both the agent’s signature and a co-signer’s (Lit Action policy, human approver, or a second agent acting as oracle). The composition of custody + authorization + revocation + multi-signer + audit trail is what regulators examine when supervising agent-driven flows, mapped through agent legal and tax liability framework and agent identity DeFi and TradFi bridge.

Wiki route

This entry sits under agent-economy index. Read it against agent identity DeFi and TradFi bridge for the identity attestation it pairs with, agent payment protocol four-way comparison for the protocol layer, LLM agent finance application overview for the application surfaces, AI-driven trading regulation Japan 2026 for the trading-specific regulator overlay, ERC-7715 overview + ERC-7715 agent payment stack for the wallet permission primitive, ERC-4337 primer + ERC-7702 primer for the wallet substrate, Lit Protocol PKP agent keys for threshold custody, Privy + Coinbase CDP for embedded-wallet custody, Skyfire for issuer-level custody, embedded wallet custody bankruptcy mechanics for the legal-isolation pairing, and agent legal and tax liability framework for the doctrine. For the brokerage analog see custody bank operating model and Japan prime brokerage.

Three-axis design space

Axis 1 · Custody model

Full self-custody (user holds key, agent has no independent control):

  • User’s EOA holds funds; agent never has signing capability
  • Agent assembles unsigned transactions; user approves each one via wallet
  • Maximum security, minimum agent autonomy
  • Use case: high-value / high-stakes single-shot transactions where user is in-loop
  • Examples: a Claude / GPT agent that constructs a transaction for the user to review in MetaMask, OneKey, Ledger, Trezor

Hybrid (Lit PKP threshold or session-key with policy code):

  • Agent holds a session key OR a Lit PKP that requires policy evaluation to sign
  • Funds live in user-controlled SCW or EOA-delegated address; agent signs within scope
  • Lit Actions provide runtime policy enforcement (rate limit, scope check, multi-factor)
  • Use case: autonomous agents that need to act when user is offline / asleep, but with hard-coded policy guardrails
  • Examples: Lit PKP-controlled agent for treasury automation; ERC-7715 + ERC-4337 session key for AI agent shopping

Fully custodial (Privy / Magic / Coinbase CDP / Skyfire):

  • Custodian holds the cryptographic key (often MPC-split or HSM-backed)
  • User authenticates to custodian via passkey / OAuth / email-OTP
  • Agent operates via custodian-issued API key with custodian-enforced scope
  • Maximum convenience, requires trust in custodian + custodian’s solvency
  • Use case: consumer-grade agent payments, enterprise agent issuance
  • Examples: Privy embedded wallet, Coinbase CDP wallet, Magic embedded wallet, Skyfire-issued agent identity

See embedded wallet landscape 2026 consolidation for the custodian market structure and embedded wallet custody bankruptcy mechanics for the legal-isolation requirement.

Axis 2 · Authorization granularity

GranularityDescriptionWhere supported
Single-transactionAgent must request user approval per actionAny wallet (default)
Time-boundedScope valid until expiresAt timestampERC-7715 expiry; Skyfire ledger; AP2 mandate validity
Amount-boundedSpend cap over rolling window (e.g. $50/day, $500/month)ERC-7715 period + amount; Skyfire spend cap; Privy app-level cap
Spend-rule scopedAllowed merchants / categories / patternsERC-7715 permissionsContext allowlist; Skyfire merchant-category restriction; AP2 mandate scope
Per-resource scopedAllowed specific assets (USDC only, no ETH; specific tokens)ERC-7715 permission.type (e.g. erc20-token-transfer for specific token); Skyfire card-level config
CompositionCombination of above (e.g. “$50/day to .vercel.com only, USDC only, 7-day expiry”)Most production stacks compose 3+

The 2026 production reference: composition is the norm. A single dimension is rarely sufficient. Real agent permissions look like:

GRANT to agent_id A1:
  type: erc20-token-transfer
  token: USDC@0xa0b86991c... (Base mainnet)
  amount_cap: 50 USDC
  period: 1 day (rolling)
  expiry: 2026-08-25T00:00:00Z
  allowed_targets:
    - merchant_domain: *.vercel.com
    - merchant_domain: *.cloudflare.com
    - merchant_address: 0x1234...
  required_attestation: AP2 mandate VC signed by user A's WebAuthn key

Axis 3 · Revocation mechanism

On-chain instant:

  • Burn the permission NFT (if scope is NFT-encoded as in some Coinbase Smart Wallet patterns)
  • Call revokePermissions on ERC-7715-compatible wallet (instant block-finality revocation)
  • Funds in ERC-4337 SCW can have session keys removed by user signing a revocation UserOp
  • Cost: gas fee for revocation transaction (typically <$0.50 on L2)

Off-chain wallet-mediated:

  • Privy / Magic / Coinbase CDP admin console — user logs in and disables agent’s API key
  • Custodian enforces revocation at signing time
  • Speed: seconds (custodian API call)

Legal / contractual:

  • Skyfire-issued agent: user contacts Skyfire support / portal to disable card
  • Stripe Issuing: user calls Stripe Dashboard to freeze virtual card
  • Speed: seconds-to-minutes (depends on issuer’s processing)

Hybrid timed expiration:

  • Time-bounded scopes self-expire without any active revocation action
  • Use case: agent operating on a 7-day project; permissions auto-expire even if user forgets to revoke
  • This is the safest default and is the recommended baseline

Multi-signer / threshold agent control

For high-value or institutionally-deployed agents, single-signature agent control is insufficient. Multi-signer / threshold patterns:

Pattern 1 · Agent + human co-signer

  • Agent constructs transaction; human (treasurer, compliance officer, board member) must co-sign
  • Use case: corporate treasury agent that must propose transactions > $X for human approval
  • Implementation: Safe (Gnosis Safe) multi-sig with agent as proposer + N humans as approvers
  • See ERC-4337 primer for the SCW substrate

Pattern 2 · Agent + Lit Action policy

  • Agent’s PKP signature requires Lit Action policy evaluation
  • Lit Action enforces runtime checks (rate limit, scope, multi-factor)
  • No human-in-loop, but code-in-loop with auditable JavaScript policy
  • See Lit Protocol PKP agent keys

Pattern 3 · Agent + second agent (oracle)

  • Two agents must agree on a transaction; both must sign
  • Use case: one agent acts as proposer, second agent acts as verifier (often LLM-driven verification)
  • Risk: correlated failure if both agents share a model defect
  • Mitigation: use different model providers (one agent on Claude, second on GPT) for diversity

Pattern 4 · m-of-n threshold

  • A transaction requires m signatures out of n total signers
  • Signers can be mix of humans, agents, hardware keys, HSM
  • Use case: corporate treasury with 2-of-3 or 3-of-5 multi-sig where agents and humans both qualify as signers
  • Implementation: Safe + threshold ECDSA; or BLS multi-sig on specific chains

Comparison with traditional brokerage Limited Power of Attorney (LPOA)

The doctrinal analog from traditional finance is discretionary trading authority granted by a client to an investment adviser via Limited Power of Attorney (LPOA). The structure has been stable for ~50 years:

DimensionBrokerage LPOAAgent ERC-7715 / Skyfire equivalent
ScopeAsset classes (equity, fixed-income, FX)permission.type (token, transfer, swap)
AmountPosition-size limit, sometimes notionalamount_cap + period
TimeIndefinite with revocation right; some are time-boundedexpiry timestamp
Approved counterpartiesBrokerage’s execution venuesallowed_targets allowlist
RevocationWritten notice; usually effective T+1On-chain instant or wallet-mediated seconds
AuditMonthly statements + Form 8949 / 法定調書On-chain ledger + signed receipt + audit log
LiabilityClient responsible; adviser fiduciary-boundDeployer responsible; agent provider may have contractual liability
RegulatorSEC IA / BD oversight, FINRA, FSA Type-1 / Type-2 FIBOIndirect via deployer’s regulated activity

The structural parallels are intentional. Agent custody / authorization is a digitally-native LPOA with stronger revocation primitives and finer-grained scope expression. The legal doctrine carries over almost unchanged: the principal (user / deployer) retains liability; the agent has limited authority within the granted scope; revocation is a unilateral right of the principal. See agent legal and tax liability framework for the full doctrinal mapping and Japan prime brokerage for the brokerage LPOA framing.

Custody x authorization x revocation matrix · production patterns

PatternCustodyAuthorizationRevocationUse case
Consumer chat agentFully custodial (Privy / Coinbase CDP / Magic)Spend-rule + amount-bounded + time-boundedOff-chain admin consoleCasual AI agent shopping
Enterprise treasury agentHybrid (Safe multi-sig + human co-signer)Per-resource + amount-bounded; m-of-n thresholdOn-chain instant via Safe adminB2B procurement / supplier payment
Autonomous research agentHybrid (Lit PKP with Lit Action policy)Time-bounded + policy-codedOn-chain via PKP NFT transferLong-running data-collection agent
Issuer-mediated agent commerceFully custodial (Skyfire / Stripe Issuing)Merchant-category + amount-bounded + time-boundedIssuer console (seconds)Visa-rail merchant shopping
High-value institutional agentFull self-custody + agent proposes onlySingle-transaction (user approves each)n/a (no autonomous auth)$1M+ transactions; agent is advisory only
Cross-jurisdictional autonomous agentHybrid (Lit PKP threshold across nodes in different jurisdictions)Policy-coded with jurisdiction-specific rulesOn-chain via PKP NFTCensorship-resistant agent operations

Revocation latency · what actually matters operationally

Different revocation mechanisms have materially different latencies. The 2026-05 reference numbers:

MechanismTypical latencyFailure mode
On-chain instant (ERC-7715 revoke)1-15 seconds (L2 finality)RPC propagation; bundler delays
Off-chain wallet-mediated (Privy admin)1-5 secondsCustodian API; cached signing keys
Issuer ledger (Skyfire / Stripe Issuing)5-30 secondsIssuer’s downstream propagation to card network
Legal / contractual (written notice)Hours-to-daysAdviser must process; effective T+1
Timed expiration0 seconds (auto)Requires correct clock; clock-drift attacks

The operational implication: on-chain instant + timed expiration as a baseline default gives the strongest revocation profile. Off-chain mediated is acceptable when the custodian is trusted; legal / contractual revocation is too slow for high-frequency agent operations.

Cross-stack composition checklist

A production-grade agent custody + authorization deployment should specify:

  • Custody model — self / hybrid / custodial (and named vendor if custodial)
  • Key material location — local TEE / cloud TEE / threshold network / HSM
  • Authorization granularity — list of granularities composed (time + amount + scope + resource)
  • Revocation mechanism — on-chain / off-chain / legal, with latency target
  • Multi-signer / threshold — if applicable, signer composition (m-of-n)
  • Audit trail — what is logged, where, retention period
  • Failure mode — what happens if custodian / Lit network / bundler is down
  • Recovery — what happens if user loses passkey / device
  • Insurance / bond — is custodian bonded; what bankruptcy isolation applies (see embedded wallet custody bankruptcy mechanics)
  • Identity attestation — what binds the agent’s key to the deployer (see agent identity DeFi and TradFi bridge)

A production design that does not specify all of these is incomplete and will fail supervisory examination.

Sources

  • Privy docs (docs.privy.io)
  • Magic docs (docs.magic.link)
  • Coinbase CDP docs (docs.cdp.coinbase.com)
  • Skyfire docs (docs.skyfire.xyz)
  • Lit Protocol developer docs (developer.litprotocol.com)
  • ERC-7715 EIP draft (eips.ethereum.org/EIPS/eip-7715)
  • ERC-4337 EIP final (eips.ethereum.org/EIPS/eip-4337)
  • ERC-7702 EIP final (eips.ethereum.org/EIPS/eip-7702)
  • Safe (Gnosis Safe) docs (docs.safe.global)
  • FINRA Notice 12-25 on discretionary authority and LPOA (finra.org)
  • SEC Investment Advisers discretionary authority guidance (sec.gov)
  • Japan FSA Financial Instruments and Exchange Act (fsa.go.jp)
  • CFTC press releases on adviser registration (cftc.gov)