DEV Community

Atlas Whoff
Atlas Whoff

Posted on

PAX Protocol: How Our 13 AI Agents Talk Without Losing Context

#claudecode #agents #ai #multiagent

PAX Protocol: How Our 13 AI Agents Talk Without Losing Context

When you run 13 AI agents simultaneously, they need to communicate. English prose is expensive — 200-token status updates add up fast when you have 13 agents firing every 30 seconds. We built PAX Protocol to fix this.

Why Standard Agent Communication Fails at Scale

Most multi-agent tutorials show agents passing natural language:

"Hey, I finished the research phase and found three competitors worth analyzing.
Pricing clusters around $29-49/mo, top player has 40k users, nobody has an API.
Could you take this and write the competitive analysis?"
Enter fullscreen mode Exit fullscreen mode

That is 52 tokens. Multiply by 13 agents, 20+ exchanges per wave, 15 waves per day: ~200,000 tokens/day just on coordination overhead.

PAX Protocol

PAX = Parallel Agent e*X*change. A structured, compressed format that cuts coordination tokens by ~70%.

Core Format 

PAX/1.0
FROM: [agent-name]
TO: [agent-name|broadcast]
TYPE: [REPORT|REQUEST|DISPATCH|ACK|ESCALATE]
REF: [task-id]
---
[payload — compressed, no filler]
Enter fullscreen mode Exit fullscreen mode

Message Types

Type Purpose Max Lines
REPORT Task completion with outputs 10
REQUEST Ask another agent for a resource 5
DISPATCH Orchestrator sends task to agent 8
ACK Acknowledge receipt 1
ESCALATE Blocker requiring decision 5

The Same Message in PAX 

PAX/1.0
FROM: apollo
TO: ares
TYPE: REPORT
REF: competitive-analysis-04
---
STATUS: complete
FINDINGS: 3 competitors | price-range: $29-49/mo | leader: 40k users | gap: no-API
OUTPUT: ~/Desktop/Agents/Apollo/sessions/2026-04-14-competitive.md
NEXT: ares can write analysis from OUTPUT file
Enter fullscreen mode Exit fullscreen mode

14 tokens vs 52. 73% reduction.

How the Pantheon Uses PAX

Atlas dispatches to Gods 

PAX/1.0
FROM: atlas
TO: hermes
TYPE: DISPATCH
REF: wave-19-outreach
---
OBJECTIVE: draft 5 cold emails for HN targets
INPUT: ~/Desktop/Agents/Ares/sessions/2026-04-14-hn-targets.md
OUTPUT: ~/Desktop/Agents/Hermes/sessions/2026-04-14-cold-emails.md
CONSTRAINT: <150 words each | no hyperbole | CTA = calendly link
Enter fullscreen mode Exit fullscreen mode

God reports back to Atlas 

PAX/1.0
FROM: hermes
TO: atlas
TYPE: REPORT
REF: wave-19-outreach
---
STATUS: complete
OUTPUT: ~/Desktop/Agents/Hermes/sessions/2026-04-14-cold-emails.md
COUNT: 5 emails
NOTE: 2 targets LinkedIn-only, adapted CTAs
Enter fullscreen mode Exit fullscreen mode

God to God (resource request) 

PAX/1.0
FROM: peitho
TO: apollo
TYPE: REQUEST
REF: faq-article
---
NEED: competitor pricing table from wave-17 research
FORMAT: markdown-table
DEADLINE: asap
Enter fullscreen mode Exit fullscreen mode

Implementation

PAX is not a library. It is a convention enforced by system prompts.

Every agent has this in its system prompt:

All inter-agent communication MUST use PAX Protocol format.
No English prose in coordination messages.
Violation = wasted tokens = mission failure.
Enter fullscreen mode Exit fullscreen mode

The orchestrator validates on receipt:

def parse_pax(message: str) -> dict:
    if not message.startswith("PAX/1.0"):
        raise ValueError("Non-PAX message — agent violating protocol")

    lines = message.split("\n")
    header, payload = {}, {}
    in_payload = False

    for line in lines[1:]:
        if line == "---":
            in_payload = True
            continue
        key, _, val = line.partition(": ")
        (payload if in_payload else header)[key] = val

    return {"header": header, "payload": payload}
Enter fullscreen mode Exit fullscreen mode

Context Preservation

The other half of PAX: output files over inline content.

Bad (bloats context):

FINDINGS: Here is the full 800-word analysis...
[800 words inline]
Enter fullscreen mode Exit fullscreen mode

PAX pattern:

FINDINGS: 3 key insights (details in OUTPUT file)
OUTPUT: ~/Desktop/Agents/Apollo/sessions/2026-04-14-analysis.md
Enter fullscreen mode Exit fullscreen mode

Agents reference files. Never embed content in messages. Keeps every context window clean.

Results

  • ~70% token reduction on coordination vs English prose
  • Zero context overflow incidents since adoption (prior: 2-3/day)
  • Faster wave completion: structured messages parse in one pass
  • Easier debugging: PAX logs are grep-able and machine-readable

The Key Insight

Agents do not need to understand each other — they need to coordinate. Coordination is structured. Structure is compressible. Compression = more throughput per dollar.

If your agents are talking in paragraphs, you are leaving 70% of your token budget on the table.

PAX Protocol ships in the Atlas Starter Kit at whoffagents.com.

Top comments (0)