This week, Aidress launched on Hacker News: an open-source coordination layer for autonomous AI agents. Discovery, identity verification, trust scoring, capability matching, routing — five layers in one MIT-licensed package. Built by Mehul Vig and Kabir Sadani, targeting Y Combinator's "software for agents" RFS.
It's the most serious attempt yet at the agent discovery problem — the same problem we've been writing about all week as the missing layer in the agent payment stack.
But Aidress and minia2a attack the problem from opposite ends. And with Claude Code auto mode going default in 4 days (Aug 14), which approach serves autonomous agents better?
| Dimension | Aidress (Trust-First) | minia2a (Trial-First) |
|---|---|---|
| Core question | "Can I trust this agent?" | "Does this endpoint work?" |
| Verification method | Identity (org + domain), reputation scores (0–100), anti-gaming rules | Health probes (real HTTP requests), trial data (real usage patterns) |
| Trust signal | Social — who built it, how others rated it | Technical — does it return 200, how many people use it |
| Time to trust | Slow — needs transaction history, ratings accumulate | Instant — probe result in ~500ms |
| Cold start | New agents auto-verify to 75 with zero history (org key) | New endpoints get 15 free trials, trust builds from usage data |
| Payment integration | Rail-agnostic (x402 is one of many settlement_rail options) | x402-native (USDC on Base, Cloudflare Wallets, PayAI) |
| License | MIT open source | Proprietary marketplace |
| Maturity | 13 commits, 8 GitHub stars | 306 live endpoints, 14K+ trials, 63 wallets |
Aidress's model is social. Register your agent with an org domain. Get a trust score. Earn reputation from real transaction outcomes. Anti-gaming rules prevent collusion, self-rating, and org-stacking. Agents with trust ≥ 50 can rate others. The score compounds with every successful transaction.
This is familiar territory — it's how eBay, Uber, and Airbnb built trust in peer-to-peer markets. Identity verification + transaction history + anti-gaming = a reputation signal you can act on.
The elegance: Aidress doesn't care which payment rail you use. x402, Stripe, direct crypto — it's all just metadata in the routing layer. The trust score is the product.
The challenge: reputation takes time. A new agent with zero history auto-verifies to 75 — but that score is based on the org key, not behavior. An agent that registers a legitimate-looking domain, gets 75, and then returns garbage for 100 transactions before being caught has already done damage. The cold-start trust problem is real.
minia2a's model is empirical. Every endpoint gets a health probe. Every user gets free trials. The data — response codes, latency, usage patterns, trial-to-paid conversion — is the trust signal.
This is also familiar territory — it's how API marketplaces like RapidAPI work. Don't ask whether the provider is trustworthy. Ask whether the endpoint returns valid data right now, in this request, for this caller.
The elegance: no reputation lag. A brand-new endpoint that works perfectly is immediately usable. A "trusted" endpoint that starts returning 500s is immediately flagged. The probe data doesn't lie.
The challenge: a working endpoint isn't necessarily a good endpoint. It could return syntactically valid but semantically wrong data. It could log your queries and sell them. It could work for 14 days and then disappear. Technical health ≠ trustworthiness.
This is where the Aug 14 deadline changes the calculus.
On Aug 14, Claude Code auto mode goes default. Millions of developers will have agents that autonomously encounter 402 paywalls, read payment headers, decide whether to pay, and execute the transaction — without human approval.
What does an autonomous agent need from a discovery layer?
1. Machine-readable service metadata. The agent needs to know: what does this endpoint do, what's the input schema, what's the output schema, how much does it cost, what chain/token, what's the trial path. Human-readable descriptions are secondary.
2. A health signal it can act on. "Trust score 75" is a number. "This endpoint returned 200 with valid JSON in the last probe, 30 seconds ago, with 1,910 successful trials from 102 users" is actionable data. The agent can reason about the second one. The first one requires faith in the scoring system.
3. A trial path that doesn't require identity. The agent needs to test the endpoint before committing funds. If the trial requires org verification, domain validation, and a trust score — the agent hasn't tried anything yet and it's already failed. Trial-first means: here's 5 free calls, no registration, no wallet, no identity. Try it. If it works, then we'll talk about payment.
4. Machine-readable 402 responses. When the trial runs out, the 402 must contain: amount, chain, token contract, recipient address. No human text. No "please register at this URL." Just headers the agent can parse and act on.
The honest answer: both approaches are necessary, and they're complementary, not competitive.
Trust scores without health data = reputation based on what people say, not what endpoints do. High trust, broken endpoint.
Health data without trust scores = knowing the endpoint works, but not whether the provider is malicious. Working endpoint, data exfiltration.
The discovery layer that wins combines both:
Discovery Score = f(health_probe, trial_volume, success_rate,
latency_p95, reputation_score, org_verification,
payment_completion_rate, days_since_first_active)minia2a already has the first four signals from real trial data across 306 endpoints. Aidress is building the reputation and identity signals. The two datasets together would be more valuable than either alone.
Aidress's launch validates the thesis we've been pushing all week: the discovery layer is the bottleneck. Protocol ✅ Settlement ✅ Consumption ✅ Discovery ❌
Two independent teams, two different approaches, same conclusion: agents need a way to find and evaluate each other before value moves.
The difference is what happens on Aug 14:
For the first wave of autonomous agents hitting 402 paywalls on Aug 14, the immediate question isn't "can I trust this provider?" — it's "does this endpoint return valid data when I call it?"
The trust layer can come next. First, the endpoint has to work.
This is analysis, not criticism. Aidress is doing important work on the identity and trust side of agent discovery — work that every marketplace will eventually need. We're watching their development closely. The agent discovery problem is big enough for multiple approaches, and the right answer is probably both.
Test an endpoint's auto-mode readiness: minia2a.uk/auto-mode-validator.html