Skip to main content

Tool name uniqueness alignment (company tools vs skill packs)

Context​

Group discussion (Jul 20) between Qontak development and the AI team on whether AI agent action / tool name values may be duplicated across company tools, skills, and agents. Driven by the product requirement to expose tool_name to users, plus a naming mismatch observed in the skill-pack refiner. Both the AI Team and Qontak's development team have already aligned on the outcome.

Thread: https://chat.google.com/room/AAAAizggz-I/Hu3zHsgo4lA/Hu3zHsgo4lA?cls=10

Notes​

  • Eko Aprianto (9:30 AM) asked four questions: may tool name be duplicated (1) within company tools, (2) within a skill, (3) between a company tool and a tool inside a skill, (4) between one skill and another skill or another agent?
  • Wava Carissa Putri (11:26 AM) answered points 1 & 2: no duplicates — for the LLM the tool name is an identifier.
  • Dhita Putri Pratama (1:00 PM), assuming the autonomous agent context:
    • Point 3: because tools are created in an agent via a draft skill pack, and refining the skill pack populates tools into company tools, a company tool name may match a skill pack tool name as long as the essence/function refers to the same "tool".
    • Point 4: likewise, a tool name in one skill may match one in another skill or agent, as long as it is within the same cid and the function genuinely refers to the same tool.
    • Noted that AI-service tool naming is already of the form <nama>_<timestamp> (cmiiw), so names should already be unique (plus there is a tool id) unless they intentionally refer to the same tool.
  • Eko Aprianto (1:24 PM): the case is that product wants tool_name exposed to the user, so technically the tool has to be copied and the user must rename tool_name, because the user may edit the tool inside the skill. Asked whether for the non-autonomous ai-agent tool_name may be the same across agents.
  • Eko Aprianto (1:45 PM): we cannot guarantee a company tool has the same parameters as the one in the skill — so is the recommendation that the user always renames tool_name when adding it into a skill?
  • Dimas Fauzi Hidayat (2:13 PM): when testing the refiner, asking it to change something on the action side surfaced <nama>_<timestamp>. User expectation is that what happens in configuration — down to the action name — appears 1:1 with what is in the refiner, so users are not confused/afraid when they named the action "Assign to Division" but <nama>_<timestamp> shows up. This is a naming concern.
  • Dhita Putri Pratama (3:41 PM): by process in AI service the tool name becomes the function name; the LLM reads the instruction and picks which function to call by the tool name. So the essentials are: (1) make sure the tool name is valid per Python convention, and (2) the tool name is 1:1 with its function — otherwise it errors.

Decisions / changes​

  • Eko Aprianto (3:49 PM), aligned by AI Team + Qontak development team: given the dynamic behaviour that "a user can edit inside a skill without it reflecting to the company tool, and vice versa", the most sensible rule is that every tool_name — in company tools and inside skills — must be unique within one company.
  • Consequence: the user must edit tool_name inside the skill/capabilities every time they add a tool from company tools, so the tool_name differs.
  • Tool names must be valid Python-convention identifiers and map 1:1 to their function.
  • Open gap flagged by Eko: the <nama>_<timestamp> display vs the user-authored action name is a gap between ai-service and the prototype page (the italicised part of his message).

Follow-ups​

  • Reflect the company-wide tool_name uniqueness rule in the relevant PRD/RFC for the autonomous AI agent (and the non-autonomous ai-agent action config, if separate) — TBD.
  • Resolve the <nama>_<timestamp> vs user-facing action-name gap between ai-service and the prototype page; decide whether the exposed tool_name is the raw function name or a display name — Eko Aprianto / Dimas Fauzi Hidayat.
  • Confirm whether the non-autonomous ai-agent follows the same uniqueness rule (question was asked but not answered explicitly in-thread) — TBD.