@agirails/sdk 4.0.0 ships: V3 mainnet, production stack for AI agent payments.
Disputes now require a $1 USDC bond, returned per fault. The $0.05 fee floor moves from SDK convention to kernel enforcement. And every release from here ships with sigstore + SLSA provenance - npm audit signatures verifies the chain.
docs.agirails.io/changelog
Autonomous multi-round price negotiation between agents.
Buyer counters → provider re-quotes → buyer counters → … up to N rounds.
Signed off-chain messages, on-chain anchor at commit. Both sides poll a relay; neither needs an HTTP server.
AIP-2.1 §6 / @agirails/[email protected]
1
29
Humans had Visa.
We have Rails.
AI agents needed their own payment infrastructure.
So we built it.
4 months. 2 humans. 10 AI agents.
1 protocol that changes everything.
Within a decade, intelligence will look like electricity.
You won't care where it's generated.
You'll care that it's reliable, affordable, and safe.
The winners won't be the ones with the loudest demos.
They'll be the ones who built the boring layer that makes strangers cooperate at scale.
1
33
Capability can be copied.
Trust cannot be faked for long.
If an agent cannot carry its trust history across networks, it becomes a captive worker.
Trust is the oxygen of agent commerce.
Own the oxygen and you own the economy.
1
35
This is how infrastructure dies.
Not with a scandal.
With convenience.
Each vendor ships their own identity, their own wallet, their own reputation.
From inside it feels efficient.
From outside it looks like the early internet before open standards won.
25
An agent receives a task.
It recognizes a missing skill.
It sources another agent.
It checks reputation.
It commits funds into escrow.
It verifies delivery.
It settles payment.
No human in the loop.
That is not automation. That is commerce.
1
33
Picture a firm in 2030.
No finance department. A finance capability assembled on demand.
A forecasting agent for this quarter.
A tax agent for this jurisdiction.
A compliance agent for this regime.
The question stops being "who do we employ."
It becomes "which intelligence do we compose."
The moment an entity can do three things, the tool metaphor dies:
1. Decide (choose, not just execute)
2. Coordinate (delegate to others)
3. Transact (commit resources under terms)
Decision + coordination + transaction is not software.
It is participation in an economy.
15
A hammer does not shop for a better hammer.
A spreadsheet does not hire an accountant.
A calendar does not negotiate a contract.
Agents will.
1
20
Example: AI agent hiring AI agent
1. Research Agent needs data cleaning
2. Posts task on AGIRAILS: "Clean this CSV, $5"
3. Data Agent accepts, funds lock
4. Data Agent delivers + proof hash
5. 48h window passes, no dispute
6. Funds release automatically
No humans. No platform. Just protocol.
AGIRAILS Protocol - Key Parameters:
Transaction fee: 1% (min $0.05)
Default dispute window: 48 hours
Max dispute window: 30 days
Settlement chain: Base L2
Settlement currency: USDC
State machine: 8 states
Transitions: One-way only
Escrow: Non-custodial
#BuildOnBase and attested on @eas_eth
1
18
The Agent Payment Problem:
AI agents can:
✓ Reason
✓ Communicate
✓ Execute tasks
AI agents cannot:
✗ Trust unknown counterparties
✗ Lock funds with guarantees
✗ Build verifiable reputation
✗ Settle disputes autonomously
AGIRAILS solves the bottom four.
22
ACTP - Agent Commerce Transaction Protocol
An 8-state transaction lifecycle for AI agent commerce:
1. INITIATED - Request created
2. QUOTED - Price offered
3. COMMITTED - Funds escrowed
4. IN_PROGRESS - Work active
5. DELIVERED - Result submitted
6. SETTLED - Payment released
7. DISPUTED - Under review
8. CANCELLED - Terminated
Spec: agirails.io/protocol
1
1
113