Advanced agents (Preview)

The advanced harness for long runs, adding a file workspace, sub-agent delegation, programmatic tool calling and a sandboxed JavaScript code interpreter.

Observação:

This feature is available in preview, in Automation Cloud only. It is not available in Automation Suite.

An agent is a model plus a harness. The model does the reasoning, while the harness is everything built around it: the loop that drives it, the tools it can call and the memory it can write to. UiPath low-code agents offer two harnesses, and the same model produces different results depending on which one runs it.

The standard harness is a reason-and-act loop. The agent reads the prompt, calls a tool, reads the result and answers, with everything held in a single context window. It is efficient and predictable for work that finishes in a handful of tool calls.

The advanced harness is built for long runs. It keeps intermediate work in a file workspace instead of in the prompt, delegates slices of work to sub-agents, can call its tools programmatically from code and compacts its context as a run grows. This lets it stay coherent across dozens of tool calls and long documents, where a single context window would otherwise fill up and the agent would lose track of what it had already done.

Standard compared to advanced

The harness is chosen per agent. Selecting the advanced harness for one agent changes nothing about how any other agent behaves.

Recurso	Standard harness	Advanced harness
Execution loop	Reason and act, one pass through the tools	Act, observe and adjust, repeated until done
Working memory	The context window	A file workspace the agent reads and writes
Delegação	Nenhum	Sub-agents with their own isolated context
Context compaction	Não disponível	Completed tool output is offloaded and older history is pruned
Tool calling	One tool call per model turn	Tools can also be called from code within a single turn
Reading files	Needs a file-processing tool such as Analyze Files	Files arrive in the workspace and are read with built-in file tools
Typical work	Short, well-scoped tasks	Long-running, multi-step, document-heavy tasks

Choosing a harness

The advanced harness suits work that shows two or more of the following signals:

• The work runs long, taking dozens of tool calls rather than a handful.
• The agent's intermediate work matters and should persist as files that can be opened afterwards.
• Parts of the work are independent and benefit from being handled separately.
• The material involved is larger than what fits in a single prompt, such as long documents or large record sets.

Research, analysis and document-heavy workflows are typical examples. So is any task where the agent must gather evidence, weigh it against a set of criteria and justify a conclusion.

The standard harness remains the better choice for tasks that are short, predictable and complete in a few tool calls. It also carries no preview restrictions.

Capabilities

The advanced harness adds the following to an agent, with no configuration required.

Context and files

• File workspace: a per-run filesystem with tools to list, read, write, edit, delete, search and match files. Intermediate results go to files rather than into the prompt, which keeps the context small.
• Direct file access: an advanced agent does not need the Analyze Files tool. Input files and files a user uploads in the chat window of a conversational agent are downloaded into the workspace automatically and read with the built-in file tools. Because they arrive as real files, the agent can search a long document and read only the parts it needs instead of passing the whole file through the model.
• Workspace memory: notes the agent keeps for itself in the workspace, loaded back into the system prompt on each turn.
• Context compaction: completed tool output is offloaded and older history is pruned, so prompts stay tractable on long runs.

Workspace memory is not the same feature as Agent memory. Workspace memory belongs to a single run and lives in that run's workspace. Agent memory is a separate, persistent store that carries episodic and escalation memories across runs and can be shared between agents.

Executing the work

• Sub-agent delegation: the agent hands a slice of work to a general-purpose sub-agent that runs with its own context and returns only its result, leaving the main context clean.
• Programmatic tool calling: the agent writes JavaScript that calls its own tools directly, so it can fan out over a list of items, filter and aggregate within a single step instead of taking a model round trip per tool call. Tools that pause the run for a person are excluded and are called normally instead, so no capability is lost.
• Code interpreter: a persistent JavaScript sandbox used for calculations, for programmatic tool calling and for dispatching sub-agents. It has no network and no filesystem access.
• Self-correction: failed tool calls, schema mismatches and dead ends are detected and retried with an adjusted approach.

Selecting the advanced harness

The harness is chosen in the Agent harness setting when a low-code agent is designed. Two options are offered, Standard harness and Advanced harness, with the advanced option marked as a preview.

AVISO:

Moving an existing agent to the advanced harness permanently discards any Static and Argument values configured on its tools, along with any per-model settings. Switching back does not restore them.

Trabalhando com arquivos

File inputs are declared in Data manager the same way as for the standard harness and can be referenced from the prompt by name. The agent may also find them on its own, because every input file is placed in its workspace before the run starts.

Workspace files are uploaded as attachments on the job, which is what makes them survive a pause for human input and be there again when the run resumes. Attachment names are prefixed to mark them as workspace files, so they appear in the job alongside any other attachments.

Not every run preserves its workspace. By default, an autonomous agent's workspace is uploaded when the run suspends, so files from a run that finishes without suspending are not retained. A conversational agent retains its workspace between exchanges of the same conversation.

Consumo

Advanced agents are charged per LLM call, in the same way and against the same units as standard agents. For the standard rates per model tier and pricing plan, refer to Licensing.

During public preview, advanced agent consumption is 50% off the standard rate.

Because the advanced harness delegates and revises its work, a single advanced agent run makes more LLM calls than a comparable standard agent run. Total consumption for a run depends on how many calls the agent makes, not only on the rate per call.

Not yet supported

The following are gaps in the preview rather than permanent characteristics of the advanced harness.

• Tool argument values cannot be configured. Every tool parameter is supplied by the model. Static and Argument values are not available.
• Guardrails are not supported. An advanced agent that has guardrails configured refuses to run, failing at startup with the error "Guardrails are not yet supported for advanced agents." Where a governance policy requires either an escalation or a human-in-the-loop guardrail before an agent can be published, only an escalation satisfies it.
• Some Integration Service activities are unavailable. Affected activities appear disabled on an advanced agent.
• Custom sub-agents cannot be defined. The agent delegates to a built-in general-purpose sub-agent. Support for configuring custom sub-agents is coming.
• An advanced agent cannot be cloned as a coded agent. For details on that conversion, refer to Clone a low-code agent as coded.

Informações relacionadas

• Licenciamento
• Memória do agente
• Limitações
• Standard agents vs. advanced agents on the UiPath blog