> ## Documentation Index > Fetch the complete documentation index at: https://docs.fetchhive.com/llms.txt > Use this file to discover all available pages before exploring further. # Creating and Configuring > Create a Hive Agent blueprint, configure stage agents, tools, repeat and compose settings, and budget guardrails Use the Hive Agent editor when you want a reusable blueprint for long-running work. The blueprint stores stage agents, tools, repeat settings, compose settings, and budget settings. Run objectives and sources are added later from **Start Run**. ## How do I create a Hive Agent? Open **Hive Agents** in the sidebar. Click **New Agent Task**. Fetch Hive creates the blueprint and opens the canvas. Rename it from the breadcrumb menu when you want a clearer name. Use the same breadcrumb menu to **Archive** a Hive Agent when you no longer need it. Archiving removes the blueprint from active Hive Agent lists and prevents new runs, but existing run records remain available in **Hive Agents** > **Logs**. {/* Add screenshot: Hive Agent editor breadcrumb menu with rename and archive actions */} ## Canvas stages Every Hive Agent runs the same loop: * **PLAN** breaks the run objective into nodes and defines what each node must produce * **EXECUTE** runs those nodes and uses the tools you attach * **VERIFY** checks each node output against the plan, source material, and verification criteria * **REPEAT** sends rejected nodes back through execute and verify until the run passes or reaches your repeat settings * **COMPOSE** assembles verified node outputs into the final response and checks that the response is faithful, complete, and in the requested format The objective is per-run. Changing the blueprint changes future runs. Existing runs keep their original objective and sources. ## How do I configure stage agents? Click **PLAN**, **EXECUTE**, **VERIFY**, or **COMPOSE** on the canvas to open that stage's settings. Each stage has a dedicated agent behind it. You can configure runtime settings such as **Charge Type**, **Model**, temperature, reasoning, prompt cache, token limits, **Tool Choice**, **Max Tool Calls**, and MCP tools. The stage instruction prompts are platform-owned. You do not edit the planner, executor, verifier, or composer instructions directly. Fetch Hive keeps those prompts fixed so the PLAN / EXECUTE / VERIFY / COMPOSE runtime stays compatible with the verifier and trace schema. New Hive Agents use these defaults unless the workspace overrides them later: | Stage | Default model | Default reasoning | | ----------- | ------------- | ----------------- | | **PLAN** | Claude Opus | High | | **EXECUTE** | Gemini Flash | Low | | **VERIFY** | Claude Opus | High | | **COMPOSE** | Claude Opus | High | If a stage model is retired, choose an active model before running the Hive Agent. ## How do tools work? Open a stage settings sheet and click **Add Tool**. Use **Tool Choice** to let the stage call tools automatically, require a tool call, or disable tool calls even when tools are attached. Use **Max Tool Calls** to cap how many tool calls that stage can make during a model turn; `0` means unlimited. Anthropic-backed stages do not show **Max Tool Calls**. Tools are most useful on **EXECUTE**, because executor nodes are where the Hive Agent gathers data, writes files, searches, or calls integrations. **PLAN**, **VERIFY**, and **COMPOSE** can also expose supported tools when the tool is compatible with that stage and model provider. Some tools have their own settings. Click an attached tool row to open its settings, such as search context size, location, or allowed domains. If you change a stage model to a provider that cannot use an attached tool, the tool remains attached and is marked **Tool not supported** until you remove it or choose a compatible model. Some tools that exist for chat Agents are hidden for Hive Agents when they do not fit the stage runtime. ## How do I configure repeat settings? Click **REPEAT** on the canvas. Set **Max repeat passes** to control how many full execute-and-verify passes the run can make. Set **Max compose attempts** to control how many times the final response can be rebuilt after all required nodes are verified. Compose retries do not send node work back through repeat. Budget is checked before each compose attempt, so conservative budget strictness can pause a run during Compose even after all required nodes are verified. Set **Completion mode**: * **Strict** fails the run if required nodes still cannot be verified * **Best effort** completes with verified work and flags anything still unverified Best effort is useful when you prefer partial verified output over a failed run after the pass limit. Best effort only applies to node verification exhaustion. If final composition cannot pass after **Max compose attempts**, the run fails. ## How do I configure budget settings? Click the dollar button in the editor control bar to open **Budget Settings**. Configure: * **Budget Hint (USD)** - soft threshold for warnings or approval prompts * **Budget Cap (USD)** - hard ceiling for the run * **When hard cap is reached** - choose **Stop run** or **Ask for approval** * **Budget Strictness** - **Conservative**, **Balanced**, or **Aggressive** * **Max Concurrent Nodes** - maximum executor nodes the planner may create and how many can run in parallel during execute/verify, from 1 to 10 Higher concurrency can finish faster and spend faster. Conservative budget strictness pauses earlier. Balanced strictness can throttle work after the hint is crossed. Aggressive strictness lets more work continue before pausing. Click the bell button in the editor control bar to open **Notifications**. Use it to choose **Send via** Email/Telegram, set an optional delivery destination, and turn on **Notify when run finishes** for future dashboard runs. {/* Add screenshot: Notifications dialog */} See also: [Running Tasks](./running-tasks), [Budget Approval](./budget-approval), and [Notifications](./notifications).