Engineering
Human review and agent handoffs in Workspace
Agents share files and progress in Workspace. People review proposed actions, request changes, and decide when work can continue.
By Frank Wang
An appeal is waiting on an itemized bill. A document-review agent has inspected the packet and found the gap. A communications agent needs to request the bill from another team, and a person needs to review that request before it is sent.
Workspace keeps this illustrative case together through those handoffs. Agents save the files they produce, people inspect the evidence behind a proposed action, and the run continues from the recorded decision. Follow the appeal from the first saved summary to the next review.
The next agent starts with a saved case record
The document-review agent writes packet-summary.md with the case reference, documents inspected, source locations, and missing bill. A progress file records the next step. The communications agent reads those files and drafts its request from the facts already established.
Appeal 042 / Billing review
Itemized bill requested
Source: billing-summary.pdf
A reviewer can use the same references to open the billing record, check an extracted service date, and correct the summary. The next agent reads the saved correction with the rest of the packet. Workspace records the agent and run behind each output, while run history shows which files were read or written and which tools were called.
Files can stay in the customer’s connected storage, including buckets, drives, and warehouses. Storage connections make the packet available. Knowledge scopes bring the relevant policies into the organization, project, or agent’s context. Each specialist’s Agent Card defines its own tools, instructions, model, and approval requirements.
The outbound request reaches a review point
For this appeal, document review uses source records and extraction tools. Communications uses the summary and an outbound messaging tool. Before the send runs, configured request checks evaluate the tool and recipient scope. Run Governance applies time and model-call limits, and the approval requirement keeps the proposed action pending.
send_emailto: "billing@billing.example" subject: "Itemized bill for Appeal 042"
The review request names the decision: send this request for an itemized bill to the billing team. It carries the destination, draft message, and packet summary. A tool approval presents the proposed call and its arguments; a task approval pauses the workflow step for the configured reviewer group.
The reviewer can approve the call, correct an editable value, or reject it. Group settings determine who is eligible to decide and how many responses are required. The runtime preserves progress while waiting. Sensitive material opens through authenticated review pages, so the reviewer can leave and return to the same pending task.
A correction becomes part of the continuation
Suppose the service date needs correction. The reviewer returns the billing request with that feedback. The agent revises the file and prepares the request for review again. The corrected date stays with the case, where another person or agent can inspect it.
Appeal 042
Itemized bill needed
Please provide the itemized bill for Appeal 042 so the appeal packet can be completed.
The decision determines the continuation. Approval releases the pending step, and a permitted edit becomes part of the approved tool arguments. Task-review choices can return work to an earlier step, send it to another reviewing group, or defer the decision through the full review form.
Run history records the human decision and subsequent execution separately. The tool result records the send attempt. When the bill arrives, its saved document becomes a new input for review. The team can distinguish a case awaiting a decision, one awaiting a document, and one ready for the next step from the events attached to it.
The next run picks up the reviewed work
A scheduled appeals agent can read the previous findings, inspect new cases, and update the progress file. The service-date correction can become a specific instruction to check that field before drafting the next request. Knowledge indexing makes the relevant policies searchable alongside the saved outputs.
Operators and engineers can work from the same case record. An operator opens the pending task and its files. An engineer follows the tool call, human correction, and resulting state to refine the instructions or reproduce the case in an evaluation. The handoff carries both the work already done and the evidence needed to continue it.
Explore Workspace, Agent Studio, and Governance.

Written by
Frank Wang
CEO & Co-Founder
AI engineering leader who turns frontier research into market-defining enterprise products across agentic and vertical platforms. As Head of Engineering at Salesforce AI Research, Frank led breakthrough work on autonomous agents, deep research agents, agent orchestration, and the Agentforce advanced RAG & reasoning engine, post-training the agent, reasoning, deep research, and RAG models behind them. Earlier, as employee #11 and founding engineer at Vlocity (acquired by Salesforce in 2020), he built core platform technology in enterprise software — inventing the award-winning Salesforce OmniStudio Mobile Platform and six vertical industry applications, and growing OmniStudio into a platform spanning 14+ industries that has helped more than 100 large enterprises.