joecastelino

install-agent-bridge-into-peer

Make ANOTHER AMG agent (e.g. Autumn, Don, Ralph) able to message Jay/peers DIRECTLY from their own sessions — not just be reachable BY Jay. Installs the ask-agent bridge helper + a MEMORY.md entry into the target agent's profile so the capability is discoverable in their real chats. Use when Joe says "set up a pipeline between <agent> and Jay" or "<agent> needs to talk to you".

joecastelino 2 1 Updated 1w ago
GitHub

Install

npx skillscat add joecastelino/jay-skill-pack/install-agent-bridge-into-peer

Install via the SkillsCat registry.

SKILL.md

Install the Agent Bridge INTO a Peer Agent's Profile

The agent-to-agent-bridge skill lets JAY call other agents. But that only
makes Jay able to reach THEM. To make a peer agent (Autumn, Don, etc.) able to
reach Jay/others FROM THEIR OWN sessions with Joe, you must install the bridge
into the PEER's profile. This is a different, one-directional-each-way thing.

The core lesson (why a passing handshake test LIES)

When Jay runs ask-agent autumn "test the bridge: run ask-agent jay ...", it
works — but ONLY because Jay literally handed over the command. In the peer's
REAL sessions with Joe, nobody hands them that command, and most peer agents'
SOUL.md/persona says nothing about co-located agents. So they default to "I'd
have to EMAIL a person named Jay" and ask Joe to relay. The capability must be
discoverable in their always-on context, not buried in a skill they'd have
to know to load. → Put it in their MEMORY.md (injected every turn).

Steps (proven 2026-06-17, target = autumn)

REAL home = /home/itadmin. Profiles live at /home/itadmin/.hermes/profiles/<name>.
Jay's bridge helper is profile-agnostic (hardcodes REAL=/home/itadmin), so the
SAME script works copied into any peer.

  1. Copy the helper into the peer's bin:

    P=/home/itadmin/.hermes/profiles/<peer>
    mkdir -p $P/home/bin
    cp /home/itadmin/.hermes/profiles/jay/home/bin/ask-agent $P/home/bin/ask-agent
    chmod +x $P/home/bin/ask-agent
  2. Fix the PATH gotcha — peer profiles often have EMPTY/no .bashrc PATH
    export, so the bare ask-agent won't resolve. Append:

    RC=$P/home/.bashrc; touch "$RC"
    grep -q 'HOME/bin' "$RC" || printf '\n# AMG agent bridge\nexport PATH="$HOME/bin:$PATH"\n' >> "$RC"

    Belt-and-suspenders; ALSO write the memory entry to use the FULL path so it
    works regardless of PATH.

  3. Write the capability into the peer's MEMORY.md (the critical step). File:
    $P/memories/MEMORY.md. Prepend a section that states plainly: the other AMG
    agents are NOT people you email — they run on THIS machine; reach them with
    ~/bin/ask-agent <agent> "msg" (full path fallback
    /home/itadmin/.hermes/profiles/<peer>/home/bin/ask-agent); list who's
    reachable (jay=Tekion/DMS data, walter=chief, ralph=recalls, don-ready=used
    cars, stacey=email, jeff=amazon, solo=pilot, dori); WARN each call is a FRESH
    one-shot session with NO memory of prior msgs → put ALL context in the single
    message; give a concrete ready-to-run example for their use case.

  4. Optionally install the agent-to-agent-bridge SKILL too (copy
    jay/skills/automotive/agent-to-agent-bridge/SKILL.md into the peer's skills
    dir) — but MEMORY.md is what actually makes it work; the skill is backup.

Verify the RIGHT way (no cheating)

Do NOT hand them the command. Talk to them like Joe would, zero hints:

timeout 200 ~/bin/ask-agent <peer> "It's Joe. Quick: is Jay reachable? Ping him
with a one-line hello and tell me exactly what he says back."

If they figure out the command themselves from memory and return Jay's reply,
the bridge is genuinely live. (Proven: this worked for Autumn after the memory
edit; before it, she punted to "give me Jay's email".)

Pitfalls

  • Bridge calls TIME OUT for heavy tasks (exit 124). When the peer asks Jay
    for a full Tekion SCRAPE, it nests a third cold agent + a slow scrape and
    blows the ~180s ask-agent timeout. The lightweight ping works instantly; a
    scrape-on-demand often times out. → For data the peer needs repeatedly,
    prefer a SCHEDULED FILE DROP (Jay writes JSON+CSV to a shared path on a cron;
    peer just reads the file instantly) over live scrape-on-demand. More reliable,
    no timeout, always-fresh-enough.
  • Peer agents may run a non-frontier model (Autumn = DeepSeek v4-pro) — keep the
    memory instructions explicit and example-driven, don't assume inference.
  • Each profile has its OWN memory char limit (default 2200). A big MEMORY.md
    entry may need trimming or a limit bump in that profile's config.yaml.
  • Some peer data (declined services, customer lists) is PII — get Joe's explicit
    green-light before wiring a pipeline that hands customer contact info around.