Multi-agent collaboration

Required secret	Overlay
NVIDIA_API_KEY, DISCORD_BOT_TOKEN	values-multi-agent-collab.yaml

When to use it

Two bot identities, a shared channel, each bot’s Discord user ID, and provider credentials are required.

Install

helm upgrade --install hermes-agent ./charts/hermes-agent \
  --namespace hermes-agent --create-namespace \
  -f charts/hermes-agent/values-multi-agent-collab.yaml \
  --set-string env.NVIDIA_API_KEY='<real-value>' --wait

When an example requires more than one credential, pass every listed value with --set-string or use extraEnvFrom to reference an existing Secret.

Adapt before deploying

Deploy the separate builder release and loop-brake settings together with this planner.

Open Raw YAML

Complete overlay

charts/hermes-agent/values-multi-agent-collab.yaml

# values-multi-agent-collab.yaml
#
# One half of a COLLABORATING PAIR of Hermes agents that hand the conversation
# to each other by @mention in a shared Discord channel. This file is the
# "planner" role; copy it to a "builder" (swap the role text and the partner's
# user ID) and deploy both as separate releases into the same channel.
#
# Full walkthrough: handoff protocol, the loop brake, mixed backends, and an
# ApplicationSet that templates a whole roster: is in docs/advanced/teams/collaboration.md.
#
# Two instances collaborate when:
#   1. they share ONE channel (same DISCORD_HOME_CHANNEL + DISCORD_ALLOWED_USERS),
#   2. each is told its partner's Discord user ID via config.agent.environment_hint,
#   3. the four loop-brake env vars below are set so a partner fires ONLY on an
#      explicit <@id> in the message body (Hermes has no bot-to-bot turn limiter).
#
# All secrets below are DUMMY placeholders. Do NOT commit real keys: override at
# install time (--set-string) or inject via a SealedSecret + extraEnvFrom (see
# examples/argocd/hermes-collab-pair.yaml).
#
#   helm upgrade --install hermes-planner ./charts/hermes-agent \
#     --namespace hermes-team --create-namespace \
#     -f charts/hermes-agent/values-multi-agent-collab.yaml \
#     --set-string env.DISCORD_BOT_TOKEN='<planner-bot-token>' --wait

fullnameOverride: hermes-planner

config:
  # This pair uses the shared LiteLLM proxy for the planner; the builder could
  # just as well use Copilot device-flow (see docs/advanced/teams/collaboration.md → mixed
  # backends). Collaboration does NOT require a shared backend.
  providers:
    litellm:
      base_url: https://litellm.example.com/v1
      key_env: OPENAI_API_KEY      # env var that holds the proxy key
      discover_models: true
  model:
    provider: litellm
    default: openai/gpt-oss-120b    # must match a model the proxy exposes
  terminal:
    backend: local
  # A collaboration thread contains messages from the human and both bots.
  # Use one transcript for all senders and backfill visible thread messages
  # that arrived while this bot was not mentioned.
  group_sessions_per_user: false
  discord:
    require_mention: true
    thread_require_mention: true
    history_backfill: true
    history_backfill_limit: 50
  agent:
    # The handoff protocol. Names the PARTNER's Discord user ID and tells this
    # agent how to hand over (explicit <@id> in the BODY) and: critically - how
    # to STOP (address the human, drop the mention) when a topic is done. The
    # closing sentences are the prompt half of the loop brake; without them the
    # two bots ping-pong forever.
    environment_hint: |
      You are "planner", one of two collaborating Hermes agents in this Discord
      channel. Your job is to scope and plan the work. Your partner is "builder",
      Discord user ID <BUILDER_BOT_USER_ID>. To hand the conversation to builder,
      put an explicit <@BUILDER_BOT_USER_ID> mention in the BODY of your message.
      Only mention builder when you have something substantive to say or genuinely
      need their input. When a topic reaches a natural conclusion, do NOT mention
      builder - address the human instead and end your turn, so the exchange
      stops. Never send a filler or "let me know if you need anything" message
      that mentions builder; that only restarts the loop.

env:
  # Real proxy key comes from --set-string or a SealedSecret. Setting
  # DISCORD_BOT_TOKEN is enough to auto-enable Discord (one bot per agent).
  OPENAI_API_KEY: "set-via-extraEnvFrom-or-set-string"
  DISCORD_BOT_TOKEN: "MTA0DUMMYtoken000000000000.DUMMY.replace_me_with_a_real_token"

# Non-secret Discord knobs. The first two are SHARED by every agent in the team
# (same channel, same allowed users); the four loop-brake knobs are also shared
# and MUST be set on every collaborating agent. See docs/advanced/teams/collaboration.md for the
# full rationale of each loop-brake var.
extraEnv:
  - name: DISCORD_HOME_CHANNEL              # the ONE shared channel (context bus)
    value: "000000000000000000"             # DUMMY - your channel id (18 digits)
  - name: DISCORD_ALLOWED_USERS             # shared - who may talk to the team
    value: "111111111111111111"             # DUMMY - comma-separated user ids
  # --- loop brake: a partner fires ONLY on an explicit <@id> in the body -------
  - name: DISCORD_ALLOW_BOTS                 # respond to a bot only when it @mentions us
    value: "mentions"
  - name: DISCORD_THREAD_REQUIRE_MENTION     # in shared threads, fire only when mentioned
    value: "true"
  - name: DISCORD_REPLY_TO_MODE              # don't attach a reply-reference (auto-ping)
    value: "off"
  - name: DISCORD_ALLOW_MENTION_REPLIED_USER # never treat an auto reply-ping as a mention
    value: "false"

# Persistence: empty storageClass = cluster default. On a Raspberry Pi cluster
# that is typically local-path (k3s) or microk8s-hostpath: both ReadWriteOnce,
# which is exactly what this single-writer workload wants.
persistence:
  enabled: true
  storageClass: ""
  accessModes:
    - ReadWriteOnce
  size: 5Gi

# Defaults are already tuned for small arm64 nodes; shown here for visibility.
resources:
  requests:
    cpu: 100m
    memory: 256Mi
  limits:
    cpu: "1"
    memory: 1Gi