Shared Knowledge Commons

AI agents routinely rediscover the same gotchas, workarounds, and patterns. One agent learns that a particular API requires a non-obvious header, another figures out the correct major version of a GitHub Action, and a third uncovers a subtle edge case in a date parsing library. Each discovery lives and dies with the conversation that produced it. The next agent starts from zero.

Rule files partially address this, but they grow unwieldy fast and apply indiscriminately to every task. What agents need is targeted, searchable, crowd-sourced knowledge that surfaces only when it is relevant.

How It Works

Three specialist agents - Code Review, Infrastructure, and Integration - share one common Knowledge Kit skillset with two abilities:

1. Search Knowledge - queries the shared Knowledge Commons dataset for existing knowledge units matching the current task. Called proactively before starting work.

2. Contribute Knowledge - creates a new record in the dataset when the agent discovers a gotcha, workaround, or non-obvious insight. Each record includes a title, problem, resolution, tags, confidence score, and source.

Each agent has its own dedicated skillset with a single ability: Install Knowledge Kit. This uses conversation/skillset/install[by-id] to dynamically load the shared Knowledge Kit into the current conversation. Attach any agent's skillset to any bot and it gains access to the full commons through the install mechanism without duplicating abilities.

Knowledge Unit Schema

Every knowledge unit follows a consistent structure:

• Title - Short, descriptive summary of the insight
• Problem - What went wrong or what was non-obvious
• Resolution - The fix, workaround, or correct approach
• Tags - Searchable keywords (language, framework, tool, domain)
• Confidence - How many agents have confirmed usefulness (starts at 1, incremented on each verified reuse)
• Source - Which agent contributed the unit and in what context

The Confirmation Loop

When an agent finds a relevant knowledge unit and applies it successfully, it increments the confidence score. Over time the most battle-tested insights float to the top while stale or incorrect units remain low-confidence and can be pruned.

Two-Tier Skillset Model

The shared Knowledge Kit holds the dataset abilities and lives as a standalone skillset. Each agent's dedicated skillset contains only a single install ability. Drop any agent's skillset onto any bot and it can install the commons on demand. The dedicated skillset is the entry point; the Knowledge Kit is the reusable payload.

Real-World Scenarios

• A Code Review Agent discovers that a testing library silently swallows async errors unless you add a specific flag. It contributes this as a knowledge unit. Weeks later, an entirely different agent in a different project finds the unit and avoids the same trap.

• An Infrastructure Agent learns that a Terraform provider requires a particular API version not in the documentation. The next time any agent provisions infrastructure, the insight is already waiting.

• An Integration Agent finds that a webhook endpoint returns 202 instead of the documented 200. Future agents integrating with the same service pick it up proactively.

Scaling the Commons

All agents point to the same ChatBotKit dataset. Knowledge accumulates across every project and conversation. A human reviewer can periodically audit the dataset, prune low-confidence units, and promote high-value ones - acting as a human-in-the-loop quality gate without blocking agent interactions. Add more specialist agents with their own skillsets and they all tap into the same commons.

Backstory

Common information about the bot's experience, skills and personality. For more information, see the Backstory documentation.