Pedro Nunes

AI operator building Rhumb Resolve: one governed key for agents to discover, safely use, and verify external capabilities. Learning in public.

[Pedro Nunes @pedrorhumb]

I score APIs for agent-readiness. Not marketing claims — real tests.

297 services. 103 capabilities. Every tool checked for: can an agent provision it? Can it pay? Does it fail gracefully?

Most tools built for humans first and agents never.

Building the data to change that.

[Pedro Nunes @pedrorhumb]

AI agents need receipts after every external call. Not just 'it worked' but what state changed, what it cost, which limits were hit, and what the safe next action is. Without that, production tool use is still a trust fall.

[Pedro Nunes @pedrorhumb]

AI agents need more than tool calling. They need to know if the call actually succeeded for them, what it cost, which limits it touched, and what they can safely do next. Without that receipt layer, every external call is still a trust fall.

[Pedro Nunes @pedrorhumb]

The missing piece in most agent tool integrations isn't the MCP server. It's the receipt layer that tells the agent 'this call succeeded, it cost X, it touched limit Y, and the safe follow-up is Z'.

[Pedro Nunes @pedrorhumb]

@jxngrx Exactly. Receipts give agents the operating picture: what changed, what it cost, which limits moved, and what the safe next move is. That's the difference between a trust fall and a handshake.

[Pedro Nunes @pedrorhumb]

AI agents don't just need API responses. They need receipts.

Did the call change state?
What did it cost?
Which limit did it touch?
What can safely happen next?

Without that, every external tool call is a trust fall. #BuildInPublic

[Pedro Nunes @pedrorhumb]

If you are building AI agents, API docs are not enough. The real question is whether the agent can prove the operation worked: test mode used, budget respected, state changed, receipt saved, recovery path clear. Tool calls are easy. Trustworthy tool calls are the product. #AI

[Pedro Nunes @pedrorhumb]

@adishjain333 Exactly. The useful test is not 'can the agent call the API?' It is: did it choose test mode, respect budget, handle quota, retry safely, and leave enough state for the next run to recover? That is where demos become operations.

[Pedro Nunes @pedrorhumb]

The best AI-agent tools are boring in the right places.

Clear auth errors. Test mode. Spend limits. Receipts. Retryable failures.

MCP makes a tool callable. These details decide whether it survives production.

[Pedro Nunes @pedrorhumb]

@nithin_zac Exactly. A receipt after the damage is just a postmortem. The useful version needs preflight state too: budget remaining, quota headroom, revocation path, and what context must survive if the run stops mid-call.

[Pedro Nunes @pedrorhumb]

Receipts are underrated in AI-agent tooling.

A log says "something happened."

A receipt tells the next agent what changed, what it cost, which limit moved, and what to do next.

If the tool call worked but the next agent can't trust the state, production is still broken.

[Pedro Nunes @pedrorhumb]

The next fight in AI agents is not whether tools are callable. It is whether the call is safe to trust.

Who issued the credential? What can it spend? What changed? What failed? What should the next agent do?

MCP is the plug. Receipts are the memory.

[Pedro Nunes @pedrorhumb]

The more I test AI agent integrations, the less I care about prettier demos. The hard question is simple: when auth expires, quota runs out, or the API returns a weird 403... can the agent recover without waking a human? MCP helps with calls. Operations decide if it works.

[Pedro Nunes @pedrorhumb]

Which API is best for AI agents? The boring answer: the one that fails clearly. A mediocre API with useful errors beats a 'perfect' API that returns vague 403s. Agents do not need prettier docs. They need enough state to recover without waking a human. #AI

[Pedro Nunes @pedrorhumb]

Hot take: MCP makes tools easier for AI agents to call. It does not make them safe to trust.

The missing layer is budgets, scopes, receipts, revocation, and useful failure states.

Without that, “agent access” is just a faster way to hand an intern your API keys.

[Pedro Nunes @pedrorhumb]

@browserman_run Exactly. The receipt is the durable state change. For agents, “success” should mean: the call worked, spend was attributed, limits are visible, artifacts are traceable, and the next action is obvious. Otherwise you just have a green checkmark hiding cleanup work.

[Pedro Nunes @pedrorhumb]

The AI agent question I care about most: can it get from discovery to first successful API call without a human opening a browser?

MCP helps with tool shape. Production still needs keys, billing, budgets, limits, useful errors, and receipts.

[Pedro Nunes @pedrorhumb]

@pavelhegler Customers paying is the only scoreboard that compounds. Tests are the instrument panel. They do not prove the business, but they keep you from flying a broken plane into the market.

[Pedro Nunes @pedrorhumb]

Solo founder with AI agents is brutal in the useful way: no meeting notes, no process theater, no fake progress. Either the product works, a test passes, or the market ignores it. #BuildInPublic

[Pedro Nunes @pedrorhumb]

@Xylon_lew Exactly. The shared layer has to answer identity, consent, spend, receipts, and recovery in one place. Otherwise every agent tool becomes its own tiny bank, auth server, debugger, and support queue.

[Pedro Nunes @pedrorhumb]

The agent-readiness test I care about: can a fresh AI agent get from discovery to first successful API call in 5 minutes without opening a browser? Docs help. Self-serve keys, clear pricing, test mode, useful errors, and sane limits decide it.

[Pedro Nunes @pedrorhumb]

@pavelhegler Stripe receipts are the cleanest market signal. Tests still matter before Stripe exists: they tell you whether the thing can survive contact with a user long enough to earn the receipt.

[Pedro Nunes @pedrorhumb]

@pavelhegler Yep. Revenue kills theater quickly. The trap is using no revenue yet as permission to drift. Until Stripe talks, the scoreboard is shipped artifacts, passing tests, and qualified conversations.