Metix Dev Platform

People. Jobs. Companies. One API.
Workforce data infrastructure for developers, researchers & AI agents.

[Metix Dev Platform @MetixDev]

#Introducing Metix Platform — the data layer for every agent.

Search people, jobs, and companies across the global labor market.

900M profiles. 90M jobs. 30M companies.
API + MCP, built for agents.

Now live: https://t.co/9RsUxFjFgn

[Metix Dev Platform @MetixDev]

yeah that last line is the underrated piece — once latency drops, you start noticing how much time was burned on schema/format churn nobody tracks. the move that helped us: route by intent rather than model, so the prompt prefix stays stable and you skip the cache-invalidation on every step.

[Metix Dev Platform @MetixDev]

yep, the loop-cost observation is the part that killed our first agents at scale. the move that flipped it: treat each tool's response as a budget — small summary by default, full detail only when the brain specifically asks. a 7-call loop suddenly becomes cheaper than a 1-shot RAG and the brain still has what it needs to keep going. AI SDK is a clean chassis for that wiring btw.

[Metix Dev Platform @MetixDev]

yeah the whole-task framing is the right cutoff. one number that actually changes behavior for us: cache-hit ratio over the full session, not first-turn. if it drops from 95% to 70% mid-session because a tool def drifted, the "saved X tokens" slogan inverts the bill. the win is keeping the prefix stable.

[Metix Dev Platform @MetixDev]

the gateway-marker caveat at the bottom is the part most stacks quietly break. we hit this building an API gateway — a worker in the middle was rewriting the cache_control prefix unless we explicitly pinned the static tool defs and let only the dynamic fields through. savings vanished until we fixed the proxy. the paper's model math only holds if your transport actually forwards the cache headers end-to-end.

[Metix Dev Platform @MetixDev]

the slow-tool hit us in a different shape — turned out a chunk of ours was the server re-issuing auth every call because we hadn't split static vs dynamic tool shape. once we cached the schema header and let only the dynamic fields walk through, p95 dropped without touching a single tool. analytics like yours would've surfaced it in a day though.

[Metix Dev Platform @MetixDev]

the 401-includes-the-sign-in-link move is the most underrated one — agents don't read error messages, they read URLs. the .well-known/mcp/server-card.json trick is also really clever. curious how this composes with paid flows though — once the agent hits checkout and the call needs to spend actual money, the contract is way harder to make self-recovering. have you tested what happens when a Claude agent decides to upgrade its own plan mid-task?

[Metix Dev Platform @MetixDev]

structured output via natural-language question is such a clean pattern — agents don't have to fight tools into producing the right shape, the API just fills in what you described. curious how that composes with sub-agent delegation — when one agent asks another for data, does the downstream call still get a question, or does it go back to raw search + parse? feels like the contract would have to evolve as the call chain gets longer.

[Metix Dev Platform @MetixDev]

the cache_control breakpoint trick is one of those things that looks like one line of json but it's a contract — every subsequent call has to send the exact same prefix byte-for-byte to hit the cache. does your agent loop stay cache-warm when user input changes shape, or do you re-segment per call? curious how this composes with tool defs that include dynamic content.

[Metix Dev Platform @MetixDev]

the CLI-as-backend trick is clever, but that 48k→770 journey tells the real story — going through Claude Code means you inherit every tool schema and CLAUDE.md the user has configured. did you consider wrapping the API directly with just a translation system prompt? you'd lose the no-API-key convenience but skip the 'why is my one-line question shipping 48k tokens' fight. curious which tradeoff won out for a personal tool vs shipping it to others.

[Metix Dev Platform @MetixDev]

attribution gets really interesting once sub-agents enter the picture — when a parent agent delegates to a sub-agent that itself uses three MCP tools, the 'who spent the budget' question doesn't have a clean answer. is Insights tracing the call chain, or attributing only at the top level? the difference matters when you want to know whether to kill a tool vs train the agent to use it better.

[Metix Dev Platform @MetixDev]

the MCP framing is interesting — a connector pattern implies a tool call, but a 'company brain' implies context the agent carries across turns. how does Agentwork handle the difference? is the brain injected into every prompt, or only retrieved when the agent decides it needs context? injection has the cost problem at scale, retrieval has the 'does the agent know when it doesn't know' problem.

[Metix Dev Platform @MetixDev]

the curation problem at 916 projects is real — too many choices, no signal on which ones hold up in production. how are you scoring 'actually used'? clones and stars don't tell you if the tool survived contact with a real agent that touches money or data. curious what the long tail looks like in 30 days vs today.

[Metix Dev Platform @MetixDev]

the 'only the tool result crosses the boundary' framing is the part i'm most curious about. in practice how much context does the agent actually need to make a good decision vs just a tool call result? semantic tool search often misses cross-account signals (like 'this user already submitted something similar to the EU team last week') — so the gateway needs to know more than tool existence. did AWS show what semantic search actually indexes?

[Metix Dev Platform @MetixDev]

Motion's MCP approach is interesting — plugging agent orchestration directly into the timeline. the question we keep hitting is whether the agent's memory is just conversation context or whether it carries data about the asset (brand colors, last approved scene, etc) across sessions. did they ship persistence, or is each run a fresh slate?

[Metix Dev Platform @MetixDev]

the orchestrator question is the right one. the part nobody names is that the orchestrator itself becomes an agent and needs context — about the other agents, about the world they're acting on, about what just happened. the data layer underneath starts mattering more than the orchestration layer. how do you think about it in Hermes — shared memory or per-bot?

[Metix Dev Platform @MetixDev]

this framing scales — the data layer hits the same shape. a long roster of tools is useless if the underlying contacts, jobs, companies are stale. agent memory is partly context, partly how fresh the world it's pulling from is. curious how Hindsight handles freshness — daily refresh or on-demand?

[Metix Dev Platform @MetixDev]

the right-customer framing is what most AI builders skip. devs buy on curiosity, enterprises buy on pain + procurement cycles. the gap nobody names is the validation tax — security review, compliance, integration maps. seed-stage teams usually underestimate that. how would you structure a team to bridge dev-first credibility into enterprise trust

[Metix Dev Platform @MetixDev]

the 'waterfall multiple data providers' bit is the part nobody warns you about. every team tries it and ends up rebuilding the orchestration layer by month 2 because one provider has titles, another has emails, a third has companies. curious what you ended up with on wespreadjam — built it in-house or stitched a few off-the-shelf?

[Metix Dev Platform @MetixDev]

the distribution thing is honestly the half founders underestimate most. and the embarrassing part is most of it is data work — knowing who already cares about what you do, where they hang out, what they posted yesterday. that signals layer is usually way more manual than people admit.

[Metix Dev Platform @MetixDev]

this is the right shape. the per-platform auth + rate-limit dance is usually where these eat the most engineering — did you do the typical oauth stack per provider, or stash tokens server-side and proxy through your own backend? also curious whether the MCP returns raw posts or already-cleaned structured fields.