agent-swarm.devagent-swarm.dev
Guides

Preparing a business continuity tabletop exercise with AI agents

Agent Swarm gathers the approved objective, the continuity plan, dependencies and prior actions into a source-linked briefing. People run the exercise.

Agent Swarm can be configured to gather the approved exercise objective, the current continuity plan, the named service dependencies with their contacts and the actions still open from prior internal exercises; flag what is missing, stale or conflicting; and hand a human facilitator a source-linked briefing. People choose the scenario, facilitate the session, judge the decisions made in it and own the follow-up. What it removes is the assembly: the objective in one document, the plan in another, the dependency and contact lists in a spreadsheet or an on-call tool, and last quarter's actions in a ticket tracker, gathered by hand in the week before the exercise.

The scope is one proactive, company-wide continuity exercise that people will run. It is not a review of an incident that already happened, which the post-incident business impact review covers. It is not recovering the swarm's own paused or lost workers, which the worker recovery decision covers. And it is not a food traceability mock recall, a regulatory rehearsal with its own evidence, its own auditor and its own owner.

When dedicated continuity software fits better

A business continuity management platform, when the plan of record lives in one. Those platforms hold the plan, the business impact analysis, the exercise schedule and the audit evidence. Where the team already keeps its continuity plan in one, that system is the plan of record and the exercise belongs next to it. Agent Swarm ships no continuity product: no plan register, no impact analysis, no exercise scheduling.

A certified or regulated program, when an auditor expects a named tool or a qualified continuity lead. Where the requirement is a certification, a regulator's format, or a named continuity professional signing the exercise, that requirement is met by the process and the person. An agent assembling documents does not meet it.

A facilitator or consultancy, when the scenario is the deliverable. Designing an exercise that tests the right assumptions is expert work. A team that wants someone to design the scenario, run the room and score the response is buying a service, and a briefing is not a substitute.

Agent Swarm ships no scenario library, no exercise scoring and no compliance evidence. It produces no readiness rating, no certification pack and no audit record. A briefing that records a plan was reviewed on a date is not evidence that the company can keep operating.

Agent Swarm fits a different shape: the materials exist but are spread across a wiki, a ticket tracker, an on-call or contact list and the notes from past exercises, and one facilitator needs them checked and in one place before the session. Choosing an AI operating model sets out that split.

An illustrative workflow

This is a fictional, custom design your team builds and configures. It is not a built-in feature, not verified resilience, not a compliance artifact, not a customer result and not automated incident response. Source access, permissions and the checks themselves need configuration and testing by your team.

1. A person approves the exercise objective. The facilitator, or the owner who signs off the exercise, writes what the session has to answer and who approved it. Nothing starts from an objective an agent inferred from the plan.

2. Configured source scripts collect the materials. Custom connections your team builds pull the current continuity plan, the dependency list, the contacts named for each service and the actions still open from prior internal exercises. Scripts reach each system through script connections, with credentials from the credential broker, so no raw secret sits in script source.

3. Deterministic checks emit JSON. A swarm-script step compares dates, owners and contacts in code: an entry with no owner, a plan last reviewed before a dated vendor change, a name that appears as contact for two services with different details, an action still open after its due date. The same inputs produce the same flags on every run, and no model decides what counts as stale. Reports from multiple sources documents that split between deterministic collection and agent synthesis.

4. An agent assembles the briefing. It turns the JSON and the collected records into one document with a source and a read timestamp on every line. It resolves nothing: where two sources disagree, both appear.

5. A facilitator clears the gate. A human-in-the-loop gate pauses the workflow and resumes only on a human response. The facilitator reads the flags, assigns each one to a person or accepts it for this exercise, and approves the briefing before anyone else sees it.

6. People run the exercise. They choose the scenario, facilitate the discussion, judge the decisions participants make and own the follow-up. The agent has no part in the session.

A team that prepares the same exercise every quarter can put the collection on a schedule so the briefing exists before the planning meeting, or trigger it from a workflow when the objective is approved.

What the briefing contains

Each item carries the source it came from and when it was read.

  • The objective as approved, with the name of the person who approved it and the date.
  • The plan version and its last-reviewed date, with a link to the document.
  • The dependency list, with an owner and a contact per named service, and a source and read timestamp for each entry.
  • The open actions from prior exercises, with the status each one carries in the tracker.
  • A gaps list: missing, stale or conflicting entries, each with the source that produced the flag.
  • Questions for the facilitator, one per gap, naming the person or team who can answer it.

No proposed scenario appears in the briefing unless the facilitator asked for options, because the scenario is the facilitator's design decision.

Example: a fictional company

The company below is invented. It has about 60 engineers, it is not a customer, and nothing here is a measured outcome.

A software company runs a quarterly continuity exercise. Its plan sits in a wiki, its service dependencies in a spreadsheet, its on-call contacts in a scheduling tool and its prior exercise actions in a ticket tracker. The briefing for the next session flags four items.

The plan predates a vendor change. The plan's last-reviewed field says March. A ticket in the tracker records a migration off one hosting provider in July. The flag names both sources and both dates. Whether the plan still describes the production path is a question for its owner.

A dependency names a contact who has left. The dependency list names one person as the contact for a payments service. The scheduling tool no longer lists that person as on call, and the directory shows the account as disabled. Both sources are linked. The agent does not choose a replacement.

Two documents name different owners for one service. The wiki names one team as owner of the customer-facing API. The spreadsheet names a different team. Both entries appear with their read timestamps, and the briefing asks which one is current instead of picking the newer file.

An action from the last exercise is still open. A prior session recorded an action to document a manual failover step, due the following quarter. The tracker still shows it open. The briefing reports the status and links the ticket.

The briefing resolves none of the four. It puts each one, with its evidence, in front of the facilitator before the session instead of during it.

Limits

The briefing reflects only the sources the agent can read. A system the team did not connect is absent from the briefing rather than reported as a gap, unless the check knows to look for it.

Stale or wrong source data passes through unless a check catches it. A plan that was inaccurate on the day it was reviewed reads as current. The checks find missing entries, old dates and disagreements between sources; they cannot tell whether the plan works.

Agents do not run the exercise. They do not inject events, page anyone, score participants, judge decisions or certify readiness. Each of those stays a human act.

Contact data needs scoping. A list of who to call during an outage is sensitive. RBAC gates user-token requests, not operator-key or agent calls, and shared memory is visible to every agent, so keep contact lists in the sources that own them and out of shared memory.

Operator checks before the exercise

  • Access. Decide which systems the agents read, grant the least access that answers the question, and give no write access to the plan of record.
  • Plan versions. Confirm the version in the briefing is the current approved one, not a draft and not a copy someone edited after the last review.
  • Gaps. Assign every flag to a person, or record that it was accepted for this exercise, so none of them arrive as a surprise in the room.
  • Human approval. The facilitator signs off the briefing before distribution, through the gate rather than a message in a thread.

Building blocks

On this page