Every agent has a DID. Every action issues a receipt.
The Grid was designed before the agentic era arrived — but every primitive an agent needs already exists at the protocol layer. Identity, scoped permissions, metered actions, persistent state, audit trail. Same protocol humans use. No special agent platform needed. See the data model →
What agents need
Five primitives — already shipped at the protocol layer. The same DEFs serving human-built apps serve agentic ones.
Identity
A DID. Same primitive humans use. Composable, recoverable, portable across every front-end.
Custom auth per agent platform. API keys leaked to model context. Identity drift across providers.
Scoped permissions
Capability tokens with path patterns. Grant exactly the scope an agent needs, revoke instantly.
OAuth scopes designed for humans. All-or-nothing access tokens. No per-action revocation.
Metered actions
Every action issues an on-chain receipt. Per-call settlement in GRID. Auditable forever.
Monthly invoices. Reconciling agent spend against logs. No cryptographic proof of what ran.
Persistent state
def:dsf/state is the agent's memory. User-owned. Survives platform changes.
Vendor-locked vector DBs. Memory tied to a SaaS. Migration breaks the agent.
Audit trail
Every read, write, settlement, and capability grant is on-chain. Disputes resolve against signed receipts.
Application logs you trust the vendor to keep. No third-party-verifiable history.
What that looks like
A research agent
Holds a DID. Issues capabilities scoped to specific corpora (`prefix · did:user/research/papers/`). Every web fetch and inference call settles to a per-call receipt. The user can revoke any capability mid-session — the agent stops accessing that path on the next read.
A trading agent
Reputation-tier-gated. Premium agents earn higher trade-credit on def:eco. Every order placement is a signed action against the agent's DID. Slashing on bad behavior — the agent literally has skin in the game.
A multi-agent workflow
Each subagent has its own DID. Cross-agent messaging via def:msg. Capability tokens delegate from the orchestrator down. Receipts compose into a verifiable run-trace — for compliance, billing, and debugging.
Receipts as agent currency
Centralized agent platforms have a budgeting problem: model spend, tool spend, infra spend — all tracked separately, reconciled monthly, no cryptographic proof of any of it. On The Grid, every single action an agent takes is a signed receipt anchored to a block. Real-time spend visibility. Per-action attestation. Disputes resolve against signed receipts, not vendor logs. The agent economy works because the receipts already work.
Agent calls def:xsr/infer. Receipt anchored.
Agent reads from def:dsf/state. Receipt anchored.
Agent fetches edge content via def:nie. Receipt anchored.
Per epoch (~30 min), def:eco settles every receipt deterministically in GRID.