DEV Community

Daniel Feldman
Daniel Feldman

Posted on

We built a Discord management team with AI agents

Discord communities become operationally heavy surprisingly fast.

At first it's manageable.

Then suddenly:

  • support requests pile up
  • repeated questions flood channels
  • moderation becomes reactive
  • events get forgotten
  • analytics disappear
  • community management becomes full-time operational work

Most servers solve this by stacking more bots.

We wanted to experiment with something different.

A persistent multi-agent team.

At "OnBuzz Community" (https://github.com/Loxia-ai/onbuzz-community?utm_source=chatgpt.com) we've been building agent systems that can coordinate operational workflows more like real teams.

Instead of one bot, we split responsibilities

For example:

Support agent

Handles:

  • repeated questions
  • troubleshooting
  • onboarding
  • documentation retrieval
  • routing complex issues

Community manager agent

Tracks:

  • unanswered discussions
  • engagement drops
  • recurring friction
  • important conversations
  • moderation signals

Event/challenge agent

Manages:

  • events
  • reminders
  • contributor challenges
  • scheduling
  • participation tracking

Analyst agent

Monitors:

  • active users
  • retention
  • growth patterns
  • popular channels
  • community trends

Manager agent

Coordinates the others:

  • delegates tasks
  • validates outputs
  • follows up on unfinished work
  • schedules operational flows
  • creates specialized agents dynamically if needed

That last part became especially interesting.

The difference memory makes

Most bots are stateless.

The second the interaction ends, operational awareness disappears.

We wanted agents that could maintain continuity.

So we built:

  • short-term memory
  • long-term memory
  • event-based memory
  • recap systems to reduce drift during long operational loops

Agents can continuously reconnect to:

  • original objectives
  • server history
  • recurring issues
  • previous workflows
  • operational context

This changes the quality of autonomous execution dramatically.

Scheduling changed the workflow

One surprisingly useful capability:

Agents can schedule tasks for:

  • themselves
  • other agents
  • recurring operational workflows

Completely configurable.

This allows long-running operational loops without constant human supervision.

Full local history access

Another important thing for us:

Local-first interaction history.

Agents can retrieve historical context dynamically when needed instead of relying only on giant prompts.

Which makes long-term operational workflows far more stable.

What became interesting technically

The challenge stopped being:
"Can the model answer correctly?"

And became:

  • how agents coordinate
  • how context flows between them
  • how memory retrieval works
  • how long-running execution stays stable
  • how orchestration scales operationally

The system starts feeling much closer to infrastructure orchestration than chatbot workflows.

Why developers care

The interesting part isn't replacing community managers.

It's reducing operational drag.

Less:

  • repetitive coordination
  • constant follow-ups
  • context rebuilding
  • manually routing tasks

More:

  • oversight
  • direction
  • execution leverage
  • parallel operational workflows

That's the area we're currently exploring.

Would genuinely love feedback from people building multi-agent systems in production.

Repo:
https://github.com/Loxia-ai/onbuzz-community

OnBuzz app:
https://onbuzz.loxia.ai

Top comments (0)