About Services Tools Blog Contact us

Agent Procedure and Tool Discovery

Lets an assistant find the right company procedure and the local tools it may run, from indexed documents and a short description of each tool. The loader that fills the index is still a prototype and does not run.

  • Agentic Workflows
  • Professional services
  • Agent skill
  • Productised
  • Local/private
  • 2026

The problem

An assistant can answer from documents but cannot see how the company works or which local tools it may run. Opening every script to learn what it needs fills up its limited working memory. So it has no cheap way to choose a tool.

What it does

Procedures, tools and skills are meant to sit in the same index as documents. Given a goal, the assistant keeps the matching procedures, the tools whose needs are already met and the skills that say when they apply. A short description file per tool lets one tool's result feed the next. The loading step currently reports that it is not built and stops, so nothing here is a live search yet.

Outcomes

  • Once the loader is built, a goal returns a ranked list of procedures, tools and skills.
  • Tools are offered only when what they need is already to hand.
  • A two-step chain runs without the assistant reading either tool's code.
  • Tools run locally, within the assistant's permitted list.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Find a procedure and the tools

Goal

The agent reads manifests only. It does not open the tool source.

Already in hand
  • in hand Date range, 6–10 Oct 2026
  • in hand Account, Harbour Wren
  • in hand Prior note, prior-note.md
  • missing Hours file, produced by the first tool
Procedures
0
Tools
0
Skills
0
Search
Waiting for a goal

Short manifests

hours-ledger-extract python

Between the inputs and the file: keep time entries inside the date range for that account, and write one row per entry.

NameTypeRequiredRole
date_rangestringyesinput
account_codestringyesinput
hours.csvfile:csvoutput

bin/hours-ledger-extract --input-file=in.json --output-dir=out/

Exit codes the agent is allowed to trust
0 success 1 input error 2 runtime error 3 partial

Capability table

NameKindDeclared needs
Weekly account noteprocedurehours, prior note
hours-ledger-extracttooldate range, account
note-draftertoolhours file, prior note
Weekly update voiceskillgoal mentions a weekly update
image-packshottoolproduct folder
Partner standing reportproceduremonth, audience
Board pack outlineproceduremeeting date
quote-from-hourstooltime entries

How it works

  1. A loading step is meant to gather procedure documents, a tool list and skill files into one index. It is not built yet.
  2. Search that index by meaning and by type for the goal.
  3. Keep tools whose needs are already met and skills that apply.
  4. Read each tool from a short description of what it needs, what it produces and the step between.
  5. Run the chosen tool on the same computer, or pass its result to a second tool when one is not enough.

Pipeline

flowchart LR
  A[Goal] --> B[Capability search]
  B --> C[Procedure]
  B --> D[Tool manifest]
  B --> E[Skill]
  D --> F[Local run]
  F --> G[Chained output]