Step 7 of 8 Advanced 15 min

From One Tool to a Toolbox: Multi-Tool Agents Without the Multi-Agent Trap

Lesson 7 — Your agent works with one tool. Someone's about to ask for a second, then a third. Here's how to grow it into a genuine toolbox without accidentally recreating the scope-creep problem you solved back in Lesson 2.

S
s4ready.ai Team

Prerequisites

  • Lesson 6: Wiring it into Joule Studio

Your one-tool agent works, and someone’s already asking for a second capability, then a third. This is the exact moment you either build a genuine toolbox — or quietly recreate the Lesson 2 scope-creep problem, one tool call at a time.

Two different decisions — don’t confuse them

Grow the toolbox vs. the multi-agent trapAdding update_bp to an agent that already has read_bp and create_bp is growing one toolbox via MCP — usually the right move. Splitting into a second agent, connected via A2A, is a different decision that should stay rare, reserved for a genuine ownership or trust boundary.GROW THE TOOLBOXread_bp + create_bp + update_bpone agent, one job, via MCP✓ Usually the right moveTHE MULTI-AGENT TRAPA second agent, connectedvia A2A, “because it sounds advanced”✗ Only when a real ownershipor trust boundary demands it
Recall AI Foundations Lesson 10: MCP connects an agent to tools; A2A connects an agent to other agents.

The test for whether a tool belongs in this toolbox

Before adding any tool, ask: does this still fit the one-sentence job description from Lesson 2? update_bp clearly does — still “manage this Business Partner.” A tool that reads Sales Order history to judge a Business Partner’s credit risk does not — that’s a different job wearing a shared vocabulary, and exactly the “and also” that should’ve been caught in Lesson 2.

Descriptions get harder to write as the toolbox grows

With one tool, an ambiguous description barely matters. With four or five, two similarly-worded descriptions can make a model reach for the wrong one — silently, with no error telling you why. As the toolbox grows, periodically reread every description side by side: could a reasonable person confuse tool A for tool B? If yes, fix it before it becomes a real incident, not after.

Try it yourself

List two plausible next tools for your Lessons 3–4 agent. Apply the one-sentence test to each. Then write their descriptions side by side and check, honestly, whether a model could mix them up.

Where this shows up in SAP

The Business Partner tutorial is a well-grown toolbox: four tools, one job, distinct descriptions — and it never adds a fifth tool that only loosely relates, because that fifth tool belongs to a different agent.

Key takeaways

  • Adding a tool and adding an agent are different decisions — most growth should be the first kind.
  • Test every new tool against the original one-sentence job. Doesn’t fit? It belongs to a different agent.
  • Reread tool descriptions side by side as the toolbox grows — ambiguity is invisible until you look for it deliberately.

Final lesson: your agent is built, tested, grown sensibly. Before it goes anywhere near production, here’s the checklist that decides whether it’s actually ready.