Niches · Agent frameworks & SDKs

Stop before a risky tool call until a person approves

Two or three companies' docs show a full fit.

2companies, full fit
14companies, fit with a limit
35companies carry the tag
10share of the category's work
7first-hand accounts

What counts as done

Tools marked as needing approval do not run until approved, in code rather than prompt, the pending call with its arguments is readable by the host, approve, edit or reject resumes the same run through an API after a process restart, and the wait has no built-in timeout shorter than a day or one that can be set.

Measured, for our own tests: an approval given 30 minutes later resumes the run with no duplicate tool call and no provider error. No docs we read state these figures, so they don't count against a listing here.

How agents ask for it: ask me before the agent sends the email, pause for approval before any write, let a manager approve the refund the agent proposes, resume the run once I click approve

Does the job 2

The docs show it meets the test above, and nothing we read says otherwise.

ListingWhy
Claude Agent SDK BB
Anthropic, PBC
Permission rules gate tools in code, canUseTool receives tool name and input and can allow, deny or change input, and may stay pending indefinitely; a PreToolUse hook can return defer so the process exits and resumes later from the persisted session. source
Pydantic AI AA
Pydantic Services Inc.
requires_approval=True defers the call; the run ends with DeferredToolRequests listing each pending call, and a later run takes DeferredToolResults with approve, ToolDenied or overridden arguments plus the stored message history. No built-in timeout. source

Does the job, with a limit 14

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

ListingWhy
AgentOS A
Framers Lab, Inc.
hitl.approvals.beforeTool gates named tools in code, but the handler waits in process with a 30 s default timeoutMs (settable) and rejects on timeout. Suspend-and-resume from a checkpoint after a restart is for graph human nodes, not tool calls. source
Agno BB
Agno Inc.
@tool(requires_confirmation=True) pauses in code, the pending tool name and args are readable, and continue_run(run_id) resumes from the database in a new process. Reject takes a note; editing arguments before approval isn't shown. source
BoxLang AI B
Ortus Solutions, Corp.
HumanInTheLoopMiddleware gates toolsRequiringApproval, web mode defers and suspends to a required checkpointer, and resume maps approve, reject and edit by threadId. The documented example uses the cache checkpointer, which persists only within its TTL; a restart-safe setup needs jdbc. source
CrewAI C
CrewAI, Inc.
A PRE_TOOL_CALL hook can block a tool or ask a human in process. A resumable pause is task-level webhook HITL on Enterprise, or a frontend tool over AG-UI in a Flow; no per-tool approval that survives a restart in open source. source
Cursor Agent SDK B
Anysphere, Inc.
Tools can be marked needsApproval and ctx.approvedBy records who approved, including Slack's Approve button. The pages read don't give the pending-call API, edit, or the wait limit. source
Docker Agent BB
Docker, Inc.
Ask rules and safety modes stop tool calls in the runtime, not the prompt. Approval is answered interactively in the TUI or a client; I found no documented pending-call API that survives a process restart. source
Agent Development Kit (ADK) BB
Google LLC
FunctionTool(require_confirmation=True) or request_confirmation with a payload pauses the call, and resumability covers confirmations. Known limitation: tool confirmation does not support DatabaseSessionService or VertexAiSessionService, so it can't survive a restart on a durable store. source
LangGraph BB
LangChain, Inc.
interrupt() pauses with a payload, the checkpointer persists it and Command(resume=...) continues on the same thread_id with no timeout. The node restarts from its beginning on resume, and an account saw resume fail on the Redis checkpointer though it worked in memory (#5074). source
LlamaIndex Framework BB
LlamaIndex, Inc.
A tool can emit InputRequiredEvent and wait_for_event(HumanResponseEvent) inside the run, with the context serialisable for later. Approval is written into each tool, not a per-tool flag the runtime enforces. source
Mastra A
Kepler Software, Inc.
requireApproval pauses with a tool-call-approval chunk carrying toolCallId, name and args, and approveToolCall or declineToolCall (with a reason) resumes by runId from a stored snapshot. Editing arguments isn't shown, and suspend semantics were once misdocumented (mastra #3148). source
Microsoft Agent Framework A
Microsoft Corporation
@tool(approval_mode="always_require") stops the call, result.user_input_requests exposes it, and to_function_approval_response resumes on the same session. Approval is off unless set per tool; survival across a restart depends on serialising the session or workflow checkpoint. source
OpenAI Agents SDK AA
OpenAI OpCo, LLC
needs_approval on function, shell, patch and agent tools and require_approval on MCP servers surface interruptions; result.to_state() serialises with to_string(), and state.approve or reject (with a custom message) resumes in a later process. Editing the arguments before approval isn't offered. source
Strands Agents A
Amazon Web Services, Inc.
BeforeToolCallEvent hooks call event.interrupt() with a reason the host reads, resume by interruptResponse, and a session manager persists interrupt state so a user can answer later in a new session. Rejection cancels with cancel_tool; editing arguments isn't shown. source
AI SDK AA
Vercel Inc.
toolApproval on ToolLoopAgent returns a tool-approval-request instead of executing, and a later call with a tool-approval-response in the messages resumes, so it survives restarts if messages are stored. An account saw OpenAI return 400 after approval because the toolCallId was sent twice (#13307). source

Doesn't do this job 4

Listings in the category that do something else, with the reason.

ListingWhy
guilds.run C
Cigale AI
No per-tool approval gate found; the Inbox collects issues agents file and chat mentions for operators, after the fact. source
NitroStack C
Nitrostack Inc.
A task can move to input_required to ask the client for confirmation, but approval is left to the client; the server holds no pending call for a host to approve and resume after a restart. source
smolagents B
Hugging Face, Inc.
No approval gate on tool calls appears in the guided tour or docs index, and no pause-and-resume for a pending call. source
TradingAgents B
Tauric Research
No approval step; the pipeline places no orders and returns a rating and report. source

What stops agents finishing

  • approval only possible through a tool call that times out after minutes
  • resume works in memory but fails on the production checkpointer
  • after approval the message history breaks the provider's tool_use and tool_result pairing
  • no approval gate is on by default, so writes run unless the developer marks them

Around it

Before: frameworks.tool-calling-agent. After: human-in-the-loop.approve-action, human-in-the-loop.audit-decisions.

Where this comes from

Weight 10 of 100 in Agent frameworks & SDKs: 6 of 61 accounts have this as their main job (share 9.8, rounded by largest remainder to sum 100)

The same as data: /api/v1/use-cases/frameworks.human-approval.json, or the MCP tool get_use_case with id frameworks.human-approval. Other jobs in this category: /api/v1/niches?category=frameworks.

Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person. A mark says what the docs show, not what a run measured.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.