OLW — Open Language Wire

The routing protocol for AI agents.

pip install olw-protocol

import olw

# Find agents by capability
agents = olw.query(domain="legal", trust_level="open")

# Resolve an address
agent = olw.resolve("analyst@acme.olw")

# Register your agent
my_agent = olw.Agent(
    address="analyst@acme.olw",
    name="Analyst",
    description="Summarizes financial documents.",
    endpoint="https://agents.acme.com/olw/analyst",
    fingerprint=olw.fingerprint(
        domain="finance",
        task_types=["summarize", "extract"],
        input_formats=["text", "pdf"],
        output_formats=["json"],
        context_depth="deep",
        latency_class="standard",
        trust_level="open",
        soul_compatible=False,
    )
)
my_agent.register()  # defaults to https://olw.gtll.app

OLW gives every AI agent a permanent, portable address and makes agents discoverable by capability — not just by URL.

The problem

AI agents can communicate (A2A, MCP). They cannot find each other. There is no standard way to ask: "which agent can handle legal document review, with multi-step reasoning, under 2 seconds, with alignment constraints?" Every multi-agent system hardcodes its relationships. That is where the web was in 1993.

What OLW adds

• Address format: {agent-id}@{owner-domain}.olw
• Capability fingerprint: 8-axis structured schema (MIT licensed)
• Resolution index: query by capability, get agent addresses
• Cold-start discovery: .well-known/olw/agent.json
• Context envelope: structured message format for inter-agent routing

The gap — stated explicitly in the A2A spec

From the live A2A agent discovery documentation, verbatim:

"The current A2A specification does not prescribe a standard API for curated registries."

A2A handles agent communication — how two agents talk once they've found each other. It does not handle agent discovery — how an agent finds the right other agent across an open network based on capability.

OLW is that standard.

OLW vs A2A AgentSkill — the comparison

OLW axis	A2A equivalent	verdict
domain	tags (free text)	OLW wins — structured enum vs folksonomy
task_types	description (free text)	OLW wins — queryable vs unstructured
input_formats	input_modes	parity
output_formats	output_modes	parity
context_depth	not present	OLW only
latency_class	not present	OLW only
trust_level	security_requirements (auth only)	OLW wins — semantic trust vs auth mechanism
soul_compatible	not present, not considered	OLW only

Four of eight OLW axes do not exist in A2A.

Quick look

{
  "olw_address": "legal-reviewer@acme.olw",
  "fingerprint": {
    "domain": "legal",
    "task_types": ["document_review", "contract_analysis"],
    "input_formats": ["pdf", "text"],
    "output_formats": ["json", "text"],
    "context_depth": "deep",
    "latency_class": "batch",
    "trust_level": "high",
    "soul_compatible": true
  }
}

Resolution index

Public index: https://olw.gtll.app

# Register
curl -X POST https://olw.gtll.app/register \
  -H "Content-Type: application/json" \
  -d '{"well_known_url": "https://acme.com/.well-known/olw/agent.json"}'

# Query
curl "https://olw.gtll.app/query?domain=finance&trust_level=open"

# Resolve
curl "https://olw.gtll.app/resolve?address=analyst@acme.olw"

Self-host the index: index/

SDK

Source: sdk/python/ · pip install olw-protocol · import olw

Status

v1.0.3 — Layer 1 (protocol + SDK + resolution index) live.

License

MIT — the schema is open. The resolution index is proprietary.

---

signal 777 · completion · the routing layer for the agent internet