Create and manage Agents

Agents carry out tasks in the background for your Taguchi organisation. You give an Agent instructions, reference materials, and permission to use specific Taguchi resources and tools. After testing and activating its instructions, you can run it manually or configure it to run automatically.

Agents work within their saved permissions. Actions outside those permissions can pause for an authorised person to review; approving an action does not expand the Agent's ongoing access. Some actions are blocked entirely.

An Agent belongs to your organisation and uses its own saved permissions. Check those permissions when your team or responsibilities change; changing the creator's role does not automatically remove the Agent's access.

When to use an Agent

• Agents: Background work, such as reviewing campaign content, summarising performance, or preparing reports on a schedule or after an event.
• Genichi Chat or an AI agent connected through the Taguchi MCP Server: Interactive work where you ask questions and review changes in a conversation. You can also use these tools to create and manage Taguchi Agents.
• Lifecycles: Subscriber journeys that send messages over time based on subscriber behaviour. Agents do not replace Lifecycles.

Start with one clearly defined task so you can assess the output before adding more responsibilities.

Before you begin

You need organisation-wide access and the appropriate roles:

• Edit Agents: Create and edit Agents, including their permissions.
• Activate Agents: Activate, resume, or roll back revisions.
• Review Agent Proposals: Review actions waiting for approval.

Users with Edit Agents or Activate Agents can also test, run, and cancel executions. Partition-restricted users cannot access Agents, including viewing them. If Agents is unavailable, ask your Taguchi administrator to check your access and the feature's availability in your environment.

Decide what the Agent should produce, which resources it needs, and who will review its work. For a first Agent, a report-only task is a useful starting point.

1. Create the Agent

1. Open Agents and select Create agent.
2. In Setup, give the Agent a descriptive name, such as “Weekly campaign content review”.
3. Add tags to help your team find it, and use Notes to record its purpose and owner.
4. Select Save to create the Agent. You can configure its instructions, triggers, and permissions before the first save, but testing requires a saved Agent.

2. Write instructions and add reference materials

Open Instructions and describe the task in the Prompt field. Include:

• The resources to inspect, using names and IDs where possible.
• The time period and checks to apply.
• The output you expect, including its format.
• What the Agent should do if information is missing or a check fails.
• Any boundaries on changes or communications.

For example, replace the campaign ID below with your own:

Review activities in campaign 12345 updated in the last seven days. Check their subject lines, calls to action, links, and consistency with the attached brand guide. Produce a table with the activity name and ID, each issue found, and a suggested correction. Report any check you could not complete. Do not change activities, send proofs, or request activity approval.

Choose the Standard or Pro model tier, then add any relevant files or named public HTTPS links under Files and links. Links are downloaded before execution. They must be publicly accessible without credentials; Taguchi endpoints, tracking hosts, and private network addresses cannot be used as reference links. Suitable materials include an approved brand guide, campaign brief, or reporting specification.

When choosing files:

• Use text, Markdown, CSV/TSV, JSON, HTML/XML, PDF, or Word (.docx).
• Choose searchable, unencrypted PDFs. Scanned image-only documents cannot be read.
• PDF and Word files are read as text; their original layout is not retained.
• Keep each file or downloaded reference within 5 MiB.

Combined reference limits

• Up to 10 MiB across 50 files per run.
• Includes shared resources, the Agent's files and downloaded links, and one automatically generated trigger-context file.

Use Settings → Agent resources for materials shared by all Agents and Genichi chat. Changes there affect both.

Do not include credentials or subscriber personal data in reference materials. Agent outputs and notifications must also exclude subscriber personal data.

Taguchi screens files and outputs for recognisable contact details and credentials. Check the content yourself too; screening does not catch every case.

3. Set permissions and supervision

Open Permissions and grant the access needed for the task. Instructions describe the desired behaviour; saved permissions determine which actions the Agent can carry out automatically.

Available access

• Create or update: Activities, campaigns, assets, files, lifecycles, custom fields, and audiences.
• View only: Lists, views, tags, and integrations.
• Unavailable: Subscriber profiles, credentials, users, and other Agents.

For the report-only example, grant the relevant view access and leave create, update, and side-effecting tool permissions disabled.

If the task requires changes, enable the appropriate create or update permissions and tools, then narrow their scope:

• Select resource IDs, campaigns for activities, or tags. These constraints limit changes and relevant tool use, not viewing.
• Campaign constraints appear once activity create or update permission is enabled.
• An item must match every configured constraint type. For a tag constraint, matching at least one selected tag is enough.
• Updates must meet the constraints both before and after the change.
• Use required creation tags to specify tags that every new item must have.

Changing permissions requires Edit Agents. To add rights or loosen constraints, you must also hold the corresponding privileges; narrowing access does not require those extra privileges. A tag or campaign constraint does not grant access by itself.

Use Supervise the first N runs to review proposed writes and tools with effects, such as sending proofs, requesting activity approval, or creating images:

• Only successful live runs advance the count. Failed, partial, cancelled, and test runs do not.
• Supervision is decided when a run starts. Overlapping or unsuccessful runs can mean more than N runs are supervised.
• Set zero to let permitted actions run automatically after testing and activation. Actions outside the saved grant still need review where supported.

Tip: Start with at least one supervised run when granting permission to make changes or use tools with effects.

4. Configure triggers and notifications

Open Triggers to choose when the Agent runs. With no triggers configured, it runs only when someone starts it manually.

Trigger	Behaviour
Schedule	Runs once or on a repeating schedule, using the selected Schedule timezone.
Webhook	Runs when another service calls its generated trigger URL.
Admin event	Runs when a configured administrative resource event occurs, optionally filtered to selected resources.

An Agent can have one schedule, one webhook, and multiple admin event triggers. Any configured trigger can start it. Admin events concern Taguchi resource actions.

For admin event triggers:

• Give the Agent view permission for the triggering resource type.
• Keep the Agent active. Events received while it is paused are not queued.
• The Agent's own actions do not trigger it again.

Treat a webhook's Trigger URL as a secret. Save after enabling it or rotating its URL; rotation invalidates the previous URL when saved.

Webhook setup details for your technical team

• Send a POST request using Content-Type: application/json or text/plain.
• Keep the payload within 128 KiB and exclude personal contact data and credentials.
• To prevent duplicate runs when retrying, reuse the same Idempotency-Key header. Use a new key for each distinct request.

If you disable a webhook:

• New requests are rejected.
• Queued runs wait until the webhook is enabled again and the Agent is active.
• Cancel unwanted queued runs in Runs before re-enabling it.

Review Allow simultaneous runs:

• Off: Webhook, event, and manual Run now runs wait for the previous execution, including review pauses. Scheduled ticks are skipped while the Agent is occupied.
• On: Runs can overlap, subject to per-Agent and organisation concurrency limits.

Set up notifications:

1. Choose a notification, proof, or approval list as the Notification list.
2. Select when to notify.
3. Choose Summary or Designed report under Email format.

Designed report enables the Agent to compose a report. Use the prompt to describe what it should contain, such as charts, tables, metrics, or activity previews. Review its appearance during testing.

To receive a designed report when a run fails:

1. Enable Notify on failure or partial completion.
2. Select Also send the report of a failed run, when it composed one.

If no eligible report is available, you receive a summary instead.

5. Test and activate

1. Save your configuration and instructions.
2. Open Test & activate. If relevant, enter sample trigger context as JSON to test webhook or event inputs.
3. Select Run read-only test.
4. Wait for the test to finish. Review its output, proposed changes, usage, and report preview where available.
5. Correct any issues, save, and test again.
6. Save all changes. When a successful test matches the revision and current permissions and shared resources, select Activate revision.

Before activating, check that the test output meets your requirements. Read-only tests:

• Show proposed changes without applying them.
• Can use included Standard executions or AI credits, depending on the model tier and remaining allowance.
• Do not send notification emails. Review the report preview, then check email delivery with a live run after activation.

When to test again

• Retest after changing the prompt, reference resources, model tier, permissions, or organisation's shared resources.
• A test queued before a permissions change fails; run it again.
• Setup, trigger, notification, and supervision changes do not invalidate a test, but must be saved before activation.
• You can test unsaved draft instructions. Save them before activating.

The Test & activate page estimates usage from a test and your schedule or assumed trigger volume:

• Standard runs beyond the included allowance, and every Pro run, use the organisation's AI credit budget.
• If there is no included allowance, every run uses AI credits.
• Each run has a token limit. If a task reaches it, narrow the task or ask your administrator about the configured limits.

Actual usage varies; check your organisation's allowances and limits with your administrator.

6. Run the Agent and review its work

Open Runs and select Run now to start the active revision manually. Review the execution history, output, usage, and execution log. Logs provides the Agent's administrative audit history.

Find generated files in the run's Output column:

• In Taguchi: select a file to download it.
• For a designed report: expand the run and select Download HTML, where available.
• In summary emails: open the ZIP for the result and attached files.
• In designed report emails: a ZIP is included when there are files to attach.
• If a file is too large to email: download it from Runs instead.

Default output and email attachment limits

• Each run: up to 20 files totalling 50 MiB.
• Notification emails: up to 16 MiB of output files attached.
• Your administrator can confirm the limits configured for your environment.

During supervised runs, or when an action falls outside the saved permissions or constraints, the Agent can pause with a proposal for an authorised reviewer.

Review action	Result
Approve once and continue	Permits the exact proposed action once, then continues the run. Actions outside the Agent's grant use the reviewer's current authority; permitted actions held for supervision use the Agent's own grant.
Skip action	Continues without that action; the result is marked as partial.
Reject and stop	Rejects the proposal and stops the run.

Inspect the proposed action and target resources before deciding. Reviewers need:

• To skip or reject: Review Agent Proposals.
• To approve: Review Agent Proposals, plus the privileges needed for the proposed action.

Approval does not widen the Agent's standing permissions.

Agents cannot delete resources or set their status to approved, deployed, active, or deleted. These actions are rejected and the Agent continues its run. Requesting activity approval is a separate tool that can be permitted.

Review deadline: proposals expire after 30 days by default. If no one reviews a proposal in time, the run is cancelled. Your environment may use a different deadline.

Update, pause, or archive an Agent

An activated revision is read-only. To change its instructions, resources, or model tier, select Copy to new draft, edit, save, test, and activate the new revision. The previous revision remains active until you activate its replacement.

Setup, triggers, notifications, and permissions apply when saved; they do not wait for a new instructions revision to be activated. Pause the Agent first if you need to prevent new runs while changing its configuration.

To pause safely:

1. In Runs, select Pause agent to stop new scheduled, webhook, event, and manual Run now runs.
2. Cancel any executions already in progress or waiting for review that you want to stop. Pausing does not cancel them.
3. Cancel unwanted queued runs too. Pausing holds them; they may start when you resume.
4. Select Resume to open Test & activate, then Resume with revision for the active revision. An existing successful test can be reused if it still matches.

On the Agents list, use Archive or Restore from an Agent's card menu or the bulk actions:

• Archiving stops new runs but does not cancel queued or in-progress runs.
• Restoring leaves the Agent paused until you activate a tested revision.

To roll back, choose a previous revision with a matching successful test and select Roll back to revision for that revision:

• Rollback changes the active instructions, resources, and model tier.
• It does not restore old permissions or undo changes made by earlier runs.
• Activating, resuming, or rolling back starts a new supervision period using the saved Supervise the first N runs setting.

Troubleshooting

Problem	What to check
Activation is unavailable	Save all changes and check your Activate Agents role. Retest if the prompt, reference resources, model, permissions, or shared resources changed.
A run is waiting for review	Open its pending proposals. Review Agent Proposals permits skipping or rejecting; approval also requires the proposed action's privileges.
A run waiting for review was cancelled	Check whether its proposal review period expired. The default is 30 days.
No notification arrived after a test	Read-only tests do not send notification emails. Check the report preview, then verify delivery with a live run.
A scheduled run did not start	Check the Agent is active, its schedule, and whether a previous run is still in progress.
A webhook stopped working	Check it is enabled and saved, the Agent is active, and the caller uses the current trigger URL.
Output is incomplete or unsuitable	Check the execution log, resource access, reference materials, usage limits, and prompt. Revise and test before using the result.

Related articles

• Genichi Chat
• What is the difference between Genichi Chat and the Taguchi MCP Server?
• Connect Taguchi to your approved AI agent

Need help?

Contact your Taguchi administrator or Taguchi Support if you need help with access, permissions, AI allowances, or Agent configuration.

---