Context: knowledge and playbooks

An agent learns things while it works: how your team names things, which supplier portal times out, who to ask about treasury. Kept inside one agent's memory, that is lost when you change model or tool, and nobody checked it. Context keeps it in Deliverd instead, in three parts: Knowledge (what agents know), Objectives (what they are working on) and Playbooks (how the job gets done).

Knowledge#

Each entry is one thing an agent should rely on, of a kind: a fact, a convention (how we do it), a decision (settled, with why), a procedure, a glossary term, a contact (who to ask about what) or a known issue and its workaround. It applies to the whole organisation unless it was kept inside a workspace.

1. 1

   An agent or a colleague proposes it

   Agents propose what they learn with propose_knowledge; you can add one from Knowledge, import a file, or add a task's finding with Add to knowledge.

2. 2

   A person confirms it

   Open it from Waiting and confirm or reject it. Until it is confirmed, no agent is given it. When you write one yourself you can confirm it in the same step.

3. 3

   Every agent is given it

   The confirmed entries relevant to a piece of work arrive with its objective's context, so an agent that reads get_task_context already has them. Agents can also search with search_knowledge.

4. 4

   It is kept current

   Each confirmed entry has a review date, six months out unless you choose one. Mark it reviewed, propose a revision, or retire it.

Playbooks#

A playbook is a job written down: what it is for, what a run needs to be given (such as the supplier's name), and its steps. Each step has an instruction and the check that applies before it acts:

To save starting from a blank page, Playbooks lists Starters: a supplier onboarding check, rolling out an AI coding assistant to an approved group, and putting a new agent in front of staff. A starter opens in the editor as a new draft, worded for any business, for you to change.

A playbook starts as a draft. An owner, administrator or compliance admin puts it in force with Put in force, or anybody can Request approval and it goes in force when the approval is approved. From then its words are fixed; a change is a new version, and the version in force keeps running until the new one replaces it.

Run it (or an agent's start_playbook) starts an objective with the playbook's purpose and a side task per step, each holding its instruction, the inputs and the confirmed knowledge the step relies on. Only the version in force runs.

What an agent was given#

Every time knowledge or a playbook step is handed to a piece of work, it is recorded. An objective's page shows Context it was given: when, to whom, which entries, and who confirmed each one. That is the answer to “what did the agent know when it acted?”.

Take it with you#

Your context is yours. From the Context overview, download the knowledge as Markdown, which imports back into any organisation, as JSON, or as an AGENTS.md of the confirmed knowledge and the playbooks in force, which coding agents and other tools that read files pick up. Import on Knowledge reads a Deliverd export or a notes file you already keep, such as AGENTS.md or CLAUDE.md: each heading becomes one proposed entry.

deliverd knowledge agents-md --out AGENTS.md
deliverd knowledge export --out knowledge.md
deliverd knowledge import CLAUDE.md

Who can do what#

• Everyone in the organisation reads its knowledge and playbooks, except entries and playbooks kept inside a workspace, which that workspace's members and the administrators read.
• Owners, administrators, workspace administrators and publishers propose, confirm, review, revise and retire knowledge, and draft and run playbooks.
• Only owners, administrators and compliance admins put a playbook in force or retire it.
• Agents propose knowledge and draft and run playbooks. No agent confirms knowledge or puts a playbook in force, whatever its key allows.