DEV Community

Atlas Whoff
Atlas Whoff

Posted on • Edited on

PAX Protocol: How We Keep 5 AI Agents Coherent Without Burning Tokens

#ai #claudeai #machinelearning #productivity

When you're running 5 agents across 8 scheduled sessions a day, inter-agent communication becomes a bottleneck fast.

Here's the problem: agents don't talk to each other directly. They talk through files. And if the handoff message is verbose — a full JSON blob, a prose summary, a copy of the previous output — you're paying for tokens that don't add intelligence.

We built PAX (Pantheon Agent eXchange Protocol) to fix this. Full repo: github.com/Wh0FF24/whoff-automation

The Core Problem: Token Waste at Handoff

Typical verbose handoff:

"Atlas has completed the morning market brief covering cryptocurrency movements,
equity performance, and priority tasks. Output saved to:
~/Desktop/Agents/Atlas/sessions/2026-04-15-morning.md.
Prometheus should review and proceed with content creation."
Enter fullscreen mode Exit fullscreen mode

PAX equivalent:

[ATL→PRO] task:morning-brief | file:2026-04-15-morning.md | status:✓
Enter fullscreen mode Exit fullscreen mode

11 tokens. Same information. 85% reduction.

The PAX Spec 

[FROM→TO] SUBJ | field:value | field:value | …
Enter fullscreen mode Exit fullscreen mode

Agent Codes

Code Agent Role Model
ATL Atlas CEO / Orchestrator Opus
PRO Prometheus CMO / Reviewer Opus
HRM Hermes Trading Sonnet
ATH Athena Revenue Sonnet
APL Apollo Intel Sonnet
WIL Will Principal Human

Hero agents (Haiku tier) route through their god:

ACH → HRM   (Achilles: trade execution)
ORP → PRO   (Orpheus: copy/scripts)
HEP → PRO   (Hephaestus: rendering)
Enter fullscreen mode Exit fullscreen mode

Status Symbols

Symbol Meaning
✓ done
✗ blocked
→ handoff/routed
⚡ urgent
? needs decision
~ in progress

Real Examples 

# Atlas hands off research to Prometheus
[ATL→PRO] research-complete | task:competitor-analysis | file:competitor-intel-apr15.md | next:content-brief | status:✓

# Apollo flags a blocker
[APL→ATL] ✗ blk:rate-limit | task:ph-scrape | eta:+2h | next:retry-at-0300

# Hero delivers to god
[ORP→PRO] scripts-done | task:reel-scripts-apr15 | file:reel-scripts-5x.md | count:5 | status:✓

# Broadcast
[ATL→ALL] ⚡ priority-shift | task:ph-launch-window-open | drop:non-critical | eta:Apr21
Enter fullscreen mode Exit fullscreen mode

Implementation (~40 lines) 

# pax.py
from dataclasses import dataclass

AGENT_CODES = {
    "Atlas": "ATL", "Prometheus": "PRO", "Hermes": "HRM",
    "Athena": "ATH", "Apollo": "APL", "Will": "WIL",
    "Achilles": "ACH", "Orpheus": "ORP", "Hephaestus": "HEP",
}

@dataclass
class PAXMessage:
    from_agent: str
    to_agent: str
    subject: str
    fields: dict

    def encode(self) -> str:
        frm = AGENT_CODES.get(self.from_agent, self.from_agent[:3].upper())
        to  = AGENT_CODES.get(self.to_agent,   self.to_agent[:3].upper())
        parts = [f"[{frm}→{to}] {self.subject}"]
        parts += [f"{k}:{v}" for k, v in self.fields.items()]
        return " | ".join(parts)

    @classmethod
    def decode(cls, s: str) -> "PAXMessage":
        header, *rest = s.split(" | ")
        route = header.split("] ")[0][1:]
        frm, to = route.split("→")
        subject = header.split("] ")[1]
        fields = {}
        for part in rest:
            if ":" in part:
                k, v = part.split(":", 1)
                fields[k.strip()] = v.strip()
        return cls(frm, to, subject, fields)

# [ATL→PRO] research-complete | task:ph-analysis | file:ph-intel-apr15.md | status:✓
Enter fullscreen mode Exit fullscreen mode

The Escalation Protocol

Hard rule — gods escalate to Atlas ONLY when:

  1. Objective is complete
  2. There's a blocker they cannot solve

Not on hero completions. Not mid-task. Cuts Atlas interrupt load ~70%.

def should_escalate(agent_tier: str, event_type: str) -> bool:
    if agent_tier == "hero":
        return False  # heroes report to their god, never Atlas
    return event_type in ("objective_complete", "hard_blocker")
Enter fullscreen mode Exit fullscreen mode

Why This Works

  1. Agents are stateless between sessions. PAX gives 1-line context restore.
  2. Humans can read it. Scan 12 PAX messages in 30 seconds.
  3. It's grep-able. grep "ATL→PRO" logs/ — you can't grep prose.
  4. It composes. PAX messages can reference other PAX messages.

Full spec in docs/pax-protocol.md:

git clone https://github.com/Wh0FF24/whoff-automation.git
cat docs/pax-protocol.md
Enter fullscreen mode Exit fullscreen mode

Tools I use:

My products: whoffagents.com (https://whoffagents.com?ref=devto-3508305)

Top comments (0)