A Research Agent to Study NASA's Space Technology Ecosystem
Agent TechPort is an autonomous agent that generates and maintains tailored knowledge bases (KBs) grounded in NASA's public TechPort (techport.nasa.gov). Using these KBs, Agent TechPort can generate tailored briefings highlighting key research findings.
Inspired by Andrej Karpathy's LLM Knowledge Bases.
Agent TechPort is an agentic research system. Each session, the agent picks a thread, queries the relevant sources, reads documents, and integrates what it finds into a persistent knowledge base — staging hypotheses, testing them, and either promoting them to content pages or filing them as dead ends. TechPort is the spine; depending on the KB, the agent also draws on USASpending, SBIR, SEC EDGAR, NTRS, and the public web. Agentic search is the mechanism; the knowledge base is the product, compounding across sessions.
These KBs are written primarily for an LLM to read, not a human. Each one is a compact, cross-linked corpus that gives a chat model the grounded context it would otherwise lack — the projects, organizations, programs, outcomes, and caveats behind a research question. The pages are human-readable too, but their density and structure are tuned for an LLM ingesting them as context.
The default way to use one: paste the KB URL (or a specific page URL) into your chat client of choice — Claude, ChatGPT, Gemini — and ask your question. The model will fetch the pages it needs and answer against them.
Tracks impact of the Small Spacecraft Technology program (recently renamed Small Spacecraft & Distributed Systems — SSDS). Agent investigated all 111 projects. 51 organization pages, 17 maturation archetypes. Same multi-source methodology as the FO tracker (TechPort + USASpending + SBIR + SEC EDGAR + web search), plus a new source: ntrs.nasa.gov.
Tracks impact of the Flight Opportunities program portfolio. Agent has access to USASpending.gov, SBIR.gov, and SEC EDGAR, and is also given access to web search, enabling it to trace outcomes across the boundary of TechPort into the private sector and other U.S. Government agencies. All combined, this crafts comprehensive infusion/transition stories.
General-purpose KB built from NASA's public TechPort database. Portfolio-level analytics (TRL distributions, field completeness, outcome tracking rates) are grounded in TechPort API queries. Some individual project and company profiles include supplementary context from public sources.
| Date | What happened | What we learned |
|---|---|---|
| 2026-04-14 | Three MCP servers broke in one day: TechPort hung 8 hours on a large GIF (stdio deadlock), USASpending flipped an endpoint GET→POST (405s), and NTRS document-download went dark mid-session. | MCP fragility — not model quality — is the dominant risk for autonomous agents. Every MCP tool needs a wall-clock timeout and an actionable error so the agent fails fast rather than hanging. |
| 2026-04-14 | SSDS KB cost ~$24.60/session — ~4× FO and ~22× TechPort-only (per-org-page: $24 vs $5.90 vs $1.80). Opus + heavy multimodal PDF reads drove the premium; large image payloads also caused stream-idle crashes early on. | Cost = model_premium × tokens. Agents doing visual document analysis should be budgeted 5–15× a text-only agent, with small default page-render caps. |
| 2026-04-10 | TechPort-only KB found to contain data from web searches, company websites, SBIR.gov, and NTRS — violating its single-source boundary | Prompt-level source boundaries are not reliable. Agents will use available tools when primary data is thin. Enforcement must be at the tool level (--disallowed-tools) from session 1, and KBs sharing a filesystem can cross-contaminate. |
| 2026-04-06 | Switched the FO agent from Sonnet to Opus. Self-correction rate tripled (6.8% → 18.9% of commits). Opus reads 80% more context before writing and produces deeper analysis. | The model matters for KB quality, not just speed. Opus self-corrects more, reads more before writing, and produces richer output at similar token efficiency per write. Both KBs now default to Opus. |
What does it take to build a knowledge base autonomously? These numbers are extracted from session trace files.
| Knowledge Base | Sess. | Model | Pages | Words | Tokens (new / cached) |
Cost |
|---|---|---|---|---|---|---|
| SST/SSDS Infusion Tracker | 50 | Opus | 51 | 120K | 23M / 520M | ~$1,230 |
| FO Infusion & Transition Tracker | 100 | Sonnet → Opus | 165 | 331K | 14M / 477M | ~$980 |
| TechPort Knowledge Base | 100 | Sonnet | 102 | 327K | 9M / 250M | ~$110 |
| Total | 318 | 778K | 46M / 1.25B | ~$2,320 |
~97% of tokens are prompt cache reads. Each session, the agent reads its standing orders, KB index, and session log, then investigates specific topics through tool calls. Because the Claude API is stateless, every tool-use turn re-sends the full conversation history — but prompt caching means previously seen in-session content is served at a fraction of the cost. The actual "new" work each session — API responses, document reads, web search results, and the agent's written output — is roughly 3% of the headline number (~23M tokens). Briefing deck generation (interactive sessions) is not included. Estimated API cost is based on Anthropic list prices as of April 2026.