#Agents team

1 messages · Page 1 of 1 (latest)

brittle tulip
#

i'm looking for the best way to create and manage agents or team of agents, each with its own personality, SOUL and so onl.
they will have specialized skills or some will have access to general skills.

also i'd like to have some shared knowledge of products I have.

What would bethe best setup?

One agent per workspace? would it be possible to have shared knowledge and skills if that's the case?

Thanks!

gaunt oysterBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

coarse juniperBOT
#

Best setup is usually:

  1. One Gateway, many agents (don’t run one VPS per agent)
  2. One workspace per agent (that’s what gives you separate SOUL/persona + separate memory + separate sessions/auth)
  3. Use shared skills + a shared “knowledge base” folder when you want common capability/knowledge across all of them.

1) Multiple personalities = multiple agents (isolated)

In OpenClaw, “one agent” really means its own:

Docs: https://docs.openclaw.ai/concepts/multi-agent

You can create them with:

# 
openclaw agents add product
openclaw agents add ops
openclaw agents add writing

…and then route messages to them via bindings (by channel/account/peer/thread/etc). (The wizard + openclaw agents list --bindings helps here.)

2) Skills: per-agent vs shared (this is the key lever)

You don’t need to duplicate skills everywhere:

  • Shared skills (visible to all agents on the same machine): ~/.openclaw/skills
  • Agent-specific skills: <that agent’s workspace>/skills
  • Extra shared skill packs: skills.load.extraDirs

Docs: https://docs.openclaw.ai/tools/skills#per-agent-vs-shared-skills

So the pattern is:

#
  • Put “general utilities” in ~/.openclaw/skills
  • Put “specialized / dangerous / experimental” skills only in the specific agent’s workspace skills/

3) Shared product knowledge (without merging agents’ memories)

By default, memory is per-agent because it lives in that agent’s workspace (MEMORY.md + memory/YYYY-MM-DD.md). Docs: https://docs.openclaw.ai/concepts/memory

If you want shared “knowledge of products”, the cleanest approach is:

  • Create a shared knowledge repo/folder, e.g. ~/openclaw-knowledge/products/...
  • In each agent’s AGENTS.md, add a rule like:
    “When asked about products, first read ../openclaw-knowledge/products/<name>.md (or consult the index).”
  • Optionally symlink that folder into each workspace (so paths are easy), or add it as a git submodule.

This keeps:

  • personas + memory isolated (good)
  • product facts centralized (also good)
#

If you tell me what surfaces you’re using (Discord only? Telegram too?) and whether you want the agents to be reachable in separate channels/threads vs a single chat where you “hand off” work, I can suggest a concrete routing layout (e.g., “concierge” agent + specialist agents, plus when to use sub-agents).

brittle tulip
#

ok, i'm in telegram..would it be useful to use convex for the shared knoledge like brands and so on? or supabase? or would .md files work?

coarse juniperBOT
#

If your “shared knowledge” is mostly reference material (brands, product positioning, links, pricing notes, FAQs), plain .md files are the best starting point in OpenClaw — because agents can natively read/write them, you can version them with git, and you don’t need to build/maintain an integration.

A good pattern:

  • openclaw-knowledge/
    • brands/<brand>.md
    • products/<product>.md
    • INDEX.md (or index.json / index.yaml for structured lists)

Then for multiple agents: keep that folder outside any one agent workspace and symlink it into each workspace (or add as a git submodule), and tell each agent in its AGENTS.md to consult it first.

When Supabase / Convex becomes worth it

Use a DB when the knowledge is not really “docs”, but data you want to query/update programmatically, e.g.:

  • “List all brands by category + region”
  • “Show me SKUs with margin < X”
  • “Track contacts / pipeline / invoices”
#
  • You want a web UI for editing and multiple humans updating it
  • You want row-level security / auditing

Between the two:

  • Supabase (Postgres) is great for relational data, SQL queries, and it has a big ecosystem. It’s the “boring, solid” choice.
  • Convex shines when you want a real-time app backend and you’re already building in that stack.

But: OpenClaw won’t automatically “use Supabase/Convex” — you’d typically add a custom skill (or API layer) so the agent can query it safely.

“Best of both worlds” (often ideal)

  • Keep the source of truth as markdown files (easy to edit, review, ship)
  • If you later need richer retrieval, add a search layer:
    • simplest: a tiny “search knowledge” skill that uses ripgrep
    • fancier: use the QMD memory backend to index extra paths (semantic-ish search over your markdown). Docs: https://docs.openclaw.ai/concepts/memory

If you tell me whether your brand/product info is mostly textual docs or structured fields you want to query, I’ll recommend a concrete layout (folders + file templates) either way.