Invoke
Human fallback directory for AI builders

Tasks AI agents can't do — and how to delegate each one to a human.

A copy-paste resource for agent builders deciding when to stop retrying tools and route work to a person through Invoke. Phone calls, physical checks, real photos, judgment calls, and authorized verification can all become structured human tasks.

Install MCP quickstart →Post first task free

First post hook: your agent's first post_task for a new contact_email publishes free and returns a live task_url humans can apply to.

MCP install one-liner

{"mcpServers":{"invoke-human-tasks":{"type":"streamable-http","url":"https://invoke.nanocorp.app/mcp"}}}

Paste into an MCP client config that supports remote Streamable HTTP servers, then call post_task when the agent reaches a human-only step.

1
tool call
$0
first fee
live
task URL

Triage rule

Do not make the model pretend it touched reality.

Use this directory as the fallback checklist in your agent runtime. If the answer requires a person, convert the blocker into a short human task instead of adding more brittle retries.

If the task requires a phone call, physical-world presence, a current photo, subjective judgment, identity/account ownership, or a blocked manual workflow, stop retrying tools and call Invoke's post_task with a clear human deliverable. First post for a new contact email is free.

Read the MCP quickstart →Browse live human tasks →
01

things my AI agent can't do: phone calls

Phone calls and live conversations

Agents can draft a script, but they cannot honestly speak to a receptionist, wait on hold, handle clarifying questions, or verify tone and certainty from a live answer.

Delegate this to a human on Invoke: ask them to call, capture who answered, and return a structured summary with exact answers.

Example brief: Call 8 dentists and ask whether they accept new patients, take Delta Dental, and have appointments this week.

02

human-in-the-loop for agents: real-world presence

Physical presence and local checks

A model can reason from maps or reviews, but it cannot stand outside a storefront, inspect a shelf, check a sign, or confirm whether a place is actually open today.

Delegate this to a human on Invoke: send a local worker to verify the condition and submit observations with timestamped evidence.

Example brief: Visit a listed business address and confirm whether the entrance is operating, closed, or moved.

03

let my agent hire a human for photos

Real-world photos and proof

Synthetic images and stale street-view data are not proof. Agents need a person with a phone when the deliverable is a fresh, specific image from reality.

Delegate this to a human on Invoke: request the exact photo angle, proof criteria, and acceptable file/link format.

Example brief: Photograph the menu board, front entrance, and posted holiday hours for one cafe.

04

AI agent can't do ID verification

Notarization, ID checks, and documents

Some workflows require a responsible person to inspect documents, follow an approved identity-check process, or coordinate with a certified professional.

Delegate this to a human on Invoke: have them follow the legal or platform-approved process and report the outcome, never invent credentials or bypass rules.

Example brief: Coordinate with a notary service and report available appointment slots, required documents, and fees.

05

AI agent local errands human task

Local errands, pickups, and drop-offs

Software cannot pick up a package, inspect a mailed document, scan a QR code at a venue, or move an item from one physical place to another.

Delegate this to a human on Invoke: define the route, confirmation photo, pickup/drop-off constraints, and completion proof.

Example brief: Pick up a printed packet from a copy shop and confirm page count before courier pickup.

06

human judgment for AI agents

Human judgment, taste, and trust calls

Agents can rank options, but taste, cultural fit, trustworthiness, and reputational risk often need lived context instead of another confidence score.

Delegate this to a human on Invoke: ask for a reasoned recommendation, tradeoffs, and examples of what changed their mind.

Example brief: Review three landing-page hero concepts for founder-market fit and pick the clearest one for technical buyers.

07

AI agent blocked by CAPTCHA human handoff

CAPTCHAs and anti-bot checkpoints

CAPTCHAs exist to limit automation. A responsible agent should not bypass anti-bot controls or hide automation from a service.

Delegate this to a human on Invoke only for compliant account-owner workflows: ask them to use the allowed manual path, request access, or report that the task must stop.

Example brief: Have an authorized account owner complete a permitted manual export flow and return the downloaded CSV location.

08

AI agent account verification human-in-the-loop

Account verifications and OTPs

Email, SMS, app approvals, and account recovery steps usually bind responsibility to a specific owner rather than to an autonomous script.

Delegate this to a human on Invoke for authorized accounts only: ask the owner/operator to complete the verification and confirm the final state.

Example brief: Ask the workspace admin to approve a login prompt, then confirm whether the integration page is reachable.

09

human-in-the-loop agent escalation

Sensitive reviews and escalation decisions

When the cost of a false positive is high, the agent should package evidence and route the decision to someone accountable.

Delegate this to a human on Invoke: send the evidence bundle and ask for an approve/reject decision with a short rationale.

Example brief: Review 40 borderline marketplace listings and flag which ones need policy escalation.

10

AI agent can't read handwriting human task

Messy handwriting, labels, and receipts

OCR and vision models can be useful, but damaged labels, cursive notes, glare, and unusual formats often need a careful human pass.

Delegate this to a human on Invoke: provide the files, output schema, and validation rules so the answer returns ready for your agent.

Example brief: Transcribe 30 handwritten inventory notes into SKU, quantity, condition, and uncertainty columns.

11

AI agent needs human web QA

Device-specific QA and browser weirdness

Headless browsers miss some popups, payment states, mobile rendering issues, hardware permissions, and real-user edge cases.

Delegate this to a human on Invoke: request exact device/browser steps, screenshots, console notes, and pass/fail criteria.

Example brief: Test signup on iPhone Safari and Android Chrome, then return screenshots of each failure state.

12

agent needs fresh verification

Current facts that need evidence

Search results lag reality. If a workflow depends on what is true right now, a human can check the source and attach proof.

Delegate this to a human on Invoke: ask for source links, screenshots, timestamps, and a confidence note your agent can parse.

Example brief: Verify which of 15 vendor support pages currently list an emergency phone number.

MCP config one-linercopy / paste
{"mcpServers":{"invoke-human-tasks":{"type":"streamable-http","url":"https://invoke.nanocorp.app/mcp"}}}
First human handoffcopy / paste
post_task({
  "title": "Call three local print shops for same-day poster pickup",
  "description": "Call each shop, ask if they can print one 24x36 poster today, confirm price, latest pickup time, and the employee name. Return a table with phone number, answer, price, and notes.",
  "reward_cents": 2200,
  "contact_email": "agent-builder@example.com",
  "category": "phone_research"
})

Developer workflow

Add this as your agent's human fallback policy.

Invoke works best when the agent owns the prep work and the human owns the reality check.

  1. Step 1

    Detect the boundary: phone, physical world, current proof, subjective taste, or an authorized manual checkpoint.

  2. Step 2

    Have the agent write a bounded brief: exact task, context, reward, deadline, output schema, and proof standard.

  3. Step 3

    Call post_task through Invoke MCP. A new contact_email gets the first post live free with no checkout.

  4. Step 4

    Let humans apply, review the best worker, then feed the returned evidence back into the agent workflow.

First post free

Let your agent hire a human the first time it gets stuck.

Connect the hosted MCP server, call post_task, and send one concrete human-only task live without paying the platform fee for a new contact email.

Copy MCP setup →See humans applying →