AI Engineer World's Fair 2026

Agents Need Receipts, Not More Tool Calls

Loading slides…

Read the talk

Agents Need Receipts, Not More Tool Calls

Froglet’s remote and local addition demos show how agents can discover external services, request work, and retain signed records across organizational boundaries.

From a talk by Armanas Povilionis

Before you start: Familiarity with API requests, Docker, and the Model Context Protocol (MCP) will help with the demonstrations.

What makes scientific automation collaborative?

What would it take to automate scientific research that other researchers can trust and build on? At Alithea Bio, Armanas Povilionis works on immuno-oncology and immunopeptidomics, using agentic automation internally and providing it to clients. Scientific research is a compelling application for agents, but completing an individual task is only part of the problem.

Adding more tools does not supply the missing collaboration mechanism. Life sciences involves contributions from multiple participants, so a result needs a record of the steps behind it. Povilionis proposes a verifiable chain of receipts: successive records that make the work inspectable and provide a foundation for trust, repeatability, and collaboration.

Dark slide with the statement “Agents need receipts, not more tool calls,” receipts highlighted in green, and a circular speaker inset.Open full source frame
Agents need receipts, not more tool calls.

Think of an agent as a cook. Better knives, more pans, and better ovens improve the speed and quality of work inside one kitchen. Scientific collaboration is closer to running a Michelin-starred restaurant: the outcome also depends on suppliers, the quality of their produce, coordinated services, and the ability to deliver the same quality repeatedly.

You cannot bring every supplier and service into the kitchen. Repeatable collaboration depends on coordinating the supply chain, including the parts outside your direct control. Local tool improvements help, but they do not establish that coordination.

0:010:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:01 · section reference included

Give agents budgets for external work

Agents already have growing collections of tools, while useful data, specialized analytics, and compute remain distributed across organizations. Povilionis’s proposed next step is to give agents budgets they can manage across those boundaries.

A token allowance for a task is an early version of this idea. A broader service budget would let an agent discover providers, request data, negotiate execution terms, and pay for work. The cook becomes an executive chef: selecting suppliers, ordering ingredients, coordinating execution, and keeping a record of what happened. This is a vision for how agents could operate as workflow automation matures.

Froglet is the open-source protocol being built to support those interactions with external data and service providers. Its scope is discovery, transactions, and verifiable receipts.

It sits between existing systems, integrating payment rails, agent harnesses, execution environments, and transport protocols. Participants need a shared interface, but they do not need identical underlying software stacks. That distinction lets an organization expose a service without adopting every other participant’s implementation.

The site presents local installation with one command and remote use with one prompt. These are entry points into the same protocol, allowing someone to try a hosted node before running one locally.

Povilionis characterizes bespoke, closed-source scientific collaborations as projects that can take years and cost millions before a reusable workflow exists. Froglet aims to reduce the integration effort after an organization has decided that a resource can be shared. Once exposed, the resource can be discovered by an agent, which can inspect its terms, request work, and receive a receipt.

After an organization has decided a resource is shareable, Povilionis estimates that exposing it through Froglet takes a few thousand tokens and minutes. These are his estimates of integration effort, not measured benchmarks; deciding what may be shared remains a prerequisite.

3:023:23
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:02 · section reference included

One node, several roles, and a signed transaction

The website walkthrough begins with documentation intended primarily for agents, alongside explanations for people. Its central architectural point is that the Froglet node implementation supports several roles.

RoleResponsibility
ProviderOffers data or a service.
RequesterRequests work from a provider.
MarketplaceOffers marketplace services as another Froglet node.

These are roles of the same node implementation, rather than separate software products that each participant must integrate.

The protocol’s responsibilities are deliberately limited: help participants find services, establish trust, pay, and prove that work happened. It does not replace the core functions of the tools and protocols underneath it.

Creating a node generates a key pair for identity and artifact signing. During execution, artifacts are signed into a chain. The intended property is tamper evidence: the chain remains valid only while its recorded data has not been altered.

Editor’s note: Froglet’s threat model distinguishes attributable records from correct results. Receipts do not prevent a provider from returning an incorrect result, so a valid chain alone cannot establish scientific correctness or reproducibility.

Discovery and execution follow different paths:

  1. Publish availability. Providers publish their services to a marketplace using the Froglet protocol.
  2. Find a service. A requester asks a known marketplace for its available service listings.
  3. Contact the provider directly. Once the requester chooses a service, quote, deal, execution, and receipt proceed through the requester–provider interaction.

The marketplace helps the participants find each other; it is not an intermediary in the subsequent execution.

Froglet walkthrough showing three connected role nodes beside explanatory text and populated terminal output.Open full source frame
Provider, requester, and marketplace roles linked by deal, register, and discover connections.

So far, this is node-to-node communication. Agents enter through plugins in Froglet’s core that connect to execution environments and agent harnesses. Povilionis names OpenClaw, NemoClaw, and operation as an MCP server or plugin.

The intended human interface is an LLM. Whether someone wants to provide a service or request one, the agent drives the protocol interaction on their behalf.

7:037:18
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:03 · section reference included

Remote trial: submit a deal and wait for its result

The terminal demonstration proceeds through a hosted trial, a local Docker installation, and a Claude session that configures and invokes a local service. The hosted trial comes first, using a Froglet node running on a remote server.

Povilionis checks that the hosted trial responds, then creates a temporary requester token. The trial token lasts 15 minutes. He also obtains access to a provider whose service adds two numbers, and points to the different tokens as evidence of different nodes in the demonstration.

Editor’s note: The current hosted-trial guide says sessions share the node’s signing identity. Different access tokens alone do not establish different cryptographic identities.

The service call has three stages:

  1. Construct the payload. Specify the schema, target service, provider, and inputs. The inputs here are 7 and 5.
  2. Submit the request. Send the payload to the remote Froglet node and obtain a deal ID. This is a handle for following the transaction; receiving an identifier is not, by itself, verification of a signed receipt.
  3. Wait for completion. Use a loop to wait for the asynchronous execution to finish. Povilionis attributes possible delays to load on the small cloud trial node.

The remote addition service returns 12 for inputs 7 and 5. The distinction between accepting a transaction and completing its computation matters even in this small example: a deal can have an identifier before its result is available.

10:4410:59
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:44 · section reference included

Run locally and distinguish identity from role

The next demonstration uses the website’s /agent bootstrap to run Froglet locally through Docker. Povilionis starts from an empty working directory, but the installer reports that the required components are already present and confirms that the node is running. The visible setup therefore is not a fresh-install timing measurement.

The local view appears to show two Froglets. Looking at their node ID, however, reveals the same identity in both entries: one node is simultaneously acting as provider and consumer.

This makes the earlier role model concrete. A role describes what a node is doing, not necessarily a separate deployment. Povilionis then shows the Docker listing, whose address and port match the local installation.

13:3013:38
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:30 · section reference included

Claude publishes and invokes the local service

With the local node running, Povilionis opens Claude and confirms that Froglet MCP is enabled. He changes the session’s permissions to reduce interruptions during the demonstration, then constrains the task to Froglet MCP and adds: “Do not install anything.”

Instead of submitting each operation separately, he supplies the prepared steps together and asks for a specific output. Claude begins by checking what is running.

The requested workflow is:

  1. Check readiness. Confirm live status and check that the provider and runtime are healthy.
  2. Publish the service. Create an artifact from the template that adds two numbers.
  3. Discover and invoke it. Find that service locally and call it with 5 and 7.
  4. Await the result. Allow for asynchronous completion, then display the requested output.

Claude executes the commands and reports the result. The local addition service returns 12 for inputs 5 and 7. The agent has handled both sides of the workflow: publishing the provider’s service and requesting its execution.

Terminal summary showing inputs 7 and 5, succeeded deal status, sum 12, settlement details, and receipt-status flags.Open full source frame
Claude reports a successful addition call: 7 plus 5 returns 12.
14:5715:13
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:57 · section reference included

Put transaction machinery beneath the service interface

The same service boundary could expose resources more useful than addition. Povilionis proposes two applications:

  • Database access: Let an agent configure a service that provides access row by row for payment.
  • GPU computation: Leave a GPU available in the background to accept computation requests.

These are proposed applications, not paid workloads demonstrated in the recording. Editor’s note: The current local quickstart describes test/regtest payment evidence and says real money has not moved.

The agent-facing surface is the service it needs to use. Receipts, payment, execution environments, negotiation, and marketplace functions sit underneath that surface. Povilionis describes this separation as reducing the need to load all the underlying tools and protocols into the LLM’s context. The benefit is architectural; no token-saving measurement accompanies it.

This connects the arithmetic demonstration back to scientific collaboration. An agent needs to find resources outside its organization, arrange remote execution, and retain a verifiable chain of what happened. Together, those capabilities let a workflow carry its record across organizational boundaries instead of ending at a successful local tool call.

Froglet’s proposed foundation for ongoing scientific automation is that combination of discovery, execution, and receipts. The demonstrations exercise the transaction path; the larger ambition is collaborative science whose successive contributions remain inspectable.

17:3317:51
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

17:33 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…