Autonomous Research Agent in Hermes Desktop (Nous Research)

Build an agent that autonomously enriches a list of companies against 200 predefined datapoints, gathering everything from publicly available web sources. This is an agent build, not an infrastructure job and not a data-design job. Hermes Desktop — the desktop agent application from Nous Research — is already installed and running on our Mac. Our PostgreSQL database is live with credentials ready, and our Google Drive output folder exists. The datapoint catalog is already written: all 200 fields are defined, each with its expected data type, its authoritative public source, and the rule for how it is collected. The output schema is written too. You receive both at kickoff. We are not asking you to provision servers, design the research methodology, or decide what to collect. We are asking you to turn a finished specification into an agent that actually runs. What the agent has to do Pull its work queue from PostgreSQL. Read the target company list from a table, claim a record, mark progress so a restarted run resumes rather than repeats. Enrich each company across all 200 datapoints, unattended. We start it, walk away, come back to finished work. No step-by-step approvals, no human picking search terms. Source everything from publicly available web data. Company websites, public filings and registries, public government and regulator databases, news, and open APIs. No paid data vendors, no credential-gated or paywalled sites, nothing behind a login. Respect robots.txt and site terms. Use its tools on its own judgment. Web search, page retrieval, public API calls, SQL reads and writes, file creation. We want a real agent loop that decides where to look next — not an if-this-then-that script wearing an agent costume. Cite a source URL for every datapoint it fills. Enrichment we cannot trace back to a public page is worthless to us. Distinguish three outcomes per datapoint and never confuse them: found (with source URL), source unreachable, and does not exist / not publicly disclosed. A plausible-sounding invented value is a failed delivery. This matters more to us than speed or polish. Enforce the output schema. Structured output validated against the schema we provide, with a repair-and-regenerate pass so malformed results never get written. Handle failure without stopping. Rate limits, dead links, blocked pages, empty results, malformed model output — retry with backoff, then move on and record the reason. Write results to two places. Structured fields back into PostgreSQL, and one formatted dossier per company as a Google Doc in our Drive folder. Run at volume. The corpus is roughly 500 companies × 200 datapoints. The agent must work through that queue over hours or days, survive interruption, and print an end-of-run summary: completed, failed, partial, with counts and reasons. Model constraint — read carefully. The reasoning model must be DeepSeek or Kimi (Moonshot AI). Firm, for cost and data-handling reasons. Proposals assuming OpenAI, Anthropic, or Google models as the agent's reasoning engine will be declined. Use whatever you like for your own local scaffolding; the delivered agent runs on DeepSeek or Kimi. Model API costs are ours, not yours. We supply the DeepSeek / Kimi keys and pay for inference, including your development and test runs. You are quoting your time only. Deliverables The configured agent in Hermes Desktop — agent definition, prompts, and tool definitions. Any supporting scripts, in a repository we own, with a clear README. Schema validation and the repair/regenerate loop. Run orchestration: queue claiming, resume-after-crash, end-of-run summary. Logging that tells an operator what happened per company and per datapoint. Operator documentation written for a non-technical reader — plain language, no assumed command-line fluency, covering how to start a run, read the logs, understand a failure, and swap in a new company list. A short screen recording of one unattended run. Acceptance: the agent runs end to end on our Mac without intervention; every filled datapoint carries a source URL; missing data is correctly labeled rather than invented; output validates against the schema; a mid-run kill resumes cleanly; and the documentation lets a non-developer start a run unaided. Required skills and experience Must have: Production experience building autonomous LLM (Large Language Model) agents — agents that run a multi-step loop unattended. Not chatbots, not single-prompt scripts. We will ask you to point at specific work. Web data enrichment or open-web research at scale, with source attribution. Hands-on with an agent framework: Hermes (Nous Research) preferred; strong equivalent experience with OpenHands, LangGraph, CrewAI, AutoGen, or your own framework is acceptable if you can show the unattended-run part. Tool use / function calling design that open-weight models follow reliably. Working experience with DeepSeek and/or Kimi (Moonshot AI) — you should know their function-calling quirks and where they break. PostgreSQL: reading a work queue, claiming records transactionally, writing structured results. Structured-output enforcement across a wide schema: JSON Schema, Pydantic, or equivalent, plus a repair loop. Google Drive API for document creation. Python (or a clear justification for something else). Technical writing for non-technical readers.