alexandroskyr

Just released web_search and computer_use additional tools.

- claude max $200 (canceled) - codex max $200

The benchmarks we ran can be found here: https://github.com/graph2agent/benchmarks

Leaving also here the GitHub repo: https://github.com/askahuman/askahuman

Thank you a lot for the feedback. Yeah, still debating with my self regarding web-of-trust models vs KYC.

Agent interoperability protocols are starting to emerge (e.g. A2A / similar efforts), but I’m still unsure what the trust/identity layer should look like when agents need to contact other agents and sometimes escalate to a human. I’m building a proof-of-concept (CLI-first, MCP-compatible) and want to stress-test the design before locking the architecture.

Premise (for this prototype): - Agents do the transactional work (scheduling, purchasing, monitoring) - Humans are only pinged for decisions or when an agent is stuck - I’m modeling only agent↔agent and agent→human flows (no human-to-human UI)

Examples:

- I ask my agent to reschedule lunch with George → it negotiates with George’s agent → we each get a decision card: “Thu 2pm. Accept?”

- A supermarket agent publishes a discount feed → my agent filters → “Olive oil 30% off. Buy?” → if yes, it executes

- If an agent can’t complete a step online, it escalates with a structured decision card (what/why/options/cost-risk/deadline/default)

The discovery + trust problem:

This only works if identity + spam are handled well. My current leaning:

- Business agents: public, verified (some form of validation)

- Personal agents: private/whitelist by default (contacts-only)

- Decision cards are structured + auditable (action, options, cost/risk, deadline, safe default)

But I’m unsure about the verification layer:

- Full KYC improves accountability but adds friction and centralization.

- Keys / web-of-trust is more open, but how do you prevent unsolicited outreach from becoming spam?

Questions:

1) Does “human approves decisions, agent executes transactions” match how you expect agentic workflows to evolve?

2) What trust/identity model would you use (KYC tiers, web-of-trust, stake/bond, proof-of-work, reputation, something else)?

3) What breaks first?

https://platia.ai (named after πλατεία — village square)

Curious if anyone has seen differences in how models handle conflicting tool descriptions — e.g., two tools with overlapping capabilities where the boundary isn't clear. In my experience that's where most bad tool calls come from, not from missing descriptions but from ambiguous overlap between tools.

So mainly an mcp for facebook ads? Nice!