KNOWLEDGE MANAGEMENT FOR AGENTS

EXAMPLES Example 1: A structured project document. A team maintains a single file named project_conventions.md at the top of their project. It states how files are named, which tools to use, how dates are written, and which folders must never be edited by hand. Whenever any agent begins work, the harness loads this file first. As a result, a drafting agent and a review agent both follow the same naming rules without being separately told. When a new rule is added, the team edits one file, and every agent picks it up on its next run. Example 2: A skill. A finance team creates a generate_invoice skill. It contains the instructions, the required fields, the tax-calculation rules, and two example invoices, one simple and one with discounts. An accounting agent and a sales-support agent both use the same skill. Their invoices look identical in structure. When the tax rule changes, the team edits the skill once and both agents are immediately up to date. Example 3: An external scratchpad. A research agent is asked to investigate a complicated topic over many steps. It creates sources_found.md and records each source it checked, with a one-line note, such as "Report A: covers 2020 only, not useful." Halfway through, its context is compacted, and the early messages are summarized. The agent re-reads its notes file and sees that it already checked Report A, so it moves on to new sources instead of repeating work. The notes also show a human reviewer exactly what has been covered. Example 4: Avoiding stale information. The knowledge base has an entry stating that the premium plan costs $29 per month, tagged as category pricing, with a last-updated date. The staleness policy says pricing entries must be reviewed every ninety days. On day ninety-one, the system flags the entry and notifies its owner. The owner checks and finds the price has changed, updates the entry, and the system re-indexes it. Without this check, the agent might have quoted the old price to customers for months.