Working with agents

Agents are the core of Delegate. Each agent has its own configuration, memory, tools, and skills, and runs tasks on your behalf.

Who can do what

Task	User	Agent owner	Admin
Create an agent	✅	✅	✅
Chat with an agent shared with them	✅	✅	✅
Clone an agent they can see	✅	✅	✅
Configure an agent (instructions, model, tools, skills)	—	✅	✅
Export an agent	—	✅	✅
Delete an agent	—	✅	✅
Open any agent in the organization	—	—	✅

Anyone shared in as an Editor can configure the agent too — see Sharing an agent.

Cloning an agent

Cloning copies an agent's configuration — its instructions, attached skills, and tool settings — into a new agent you own. This is the fastest way to start from a known-good setup. The conversation history and private memory are not copied.

Exporting an agent

Agent → Settings → Export downloads the agent as a single portable .zip so it can be moved to another platform. You need to be the agent's owner or an admin — being shared in as an Editor is not enough, because the export hands over the agent's complete history in one file.

The bundle contains:

• Transcripts — every session that has not been deleted, with its messages in order.
• Memories — the agent's long-term memory, with each entry's type.
• Skills — every attached skill's instructions, manifest, and files.
• Model slug — the identifier of the model the agent runs on (for example anthropic/claude-opus-4), plus its fallback chain. The destination platform looks the agent's model up by this slug.
• Configuration — name, description, description prompt, initial greeting, bootstrap instructions, the names of attached tools, and budget settings.

Two things deliberately stay behind. Credentials never travel — tools are listed by name only, and no API key, OAuth token, or tool configuration is written into the bundle; you reconnect the tools on the destination. And embeddings are not exported — the destination re-computes them with its own embedding model.

If any file of any attached skill cannot be read — a missing object in storage, or a malformed file path — the export fails rather than handing back an archive that is missing it. An incomplete bundle looks importable and isn't, and you would only find out when the imported skill failed to run. The error names the file so it can be fixed or removed.

Very large agents are refused rather than truncated, for the same reason.

Each export is written to the audit log.

Configuring an agent

Agent → Settings holds everything about how an agent behaves. You need to be its owner, an Editor, or an admin.

• General — name, description, description prompt (how the agent describes itself), initial greeting, whether the agent is Active, and its Model. If your organization has brought its own key, agents can run on that key.
• Tool Bundles — related tools enabled and configured as a group.
• Tools — individual capabilities, each with its own parameters and credentials.
• Bootstrap — instructions the agent receives on its first task session, layered on top of the organization default set in Organization → General.
• Skills — reusable capability packages the agent can run.
• Memories — what the agent has learned, editable by hand.
• Sharing, Share Links, and Transfer Ownership — who else can use it.

Talking to an agent

The chat input has three modes — Listen, Plan, and Go — which change whether the agent acts immediately, proposes first, or waits for you to finish your thought. See Chatting: Listen, Plan, Go.

Past conversations, search, and resumption are covered in Sessions & history.

Scheduling work (routines)

Agents can run on a schedule using routines — a cron schedule, a fixed interval, or a one-shot at a future time. Useful for recurring reports, monitoring, and reminders.

There is no routines screen: you set one up by asking the agent, in plain language ("check the deploy queue every weekday at 9am and message me if anything is stuck"). See Scheduling with routines.

Routines respect the same budgets as interactive tasks, and their spend appears in Usage & spend like any other.

Memory

Agents keep long-term memory of facts, procedures, and lessons learned, and can search their own past sessions to recall earlier decisions and context. See Memory.

Files

Each agent has its own file area — a writable workspace, the attachments you've sent it, and the files its skills provide. See Agent files.