THIS IS THE IDEA I CAN'T STOP THINKING ABOUT
nobody shipped this. it's just my thought. but i'm pretty sure Opus 5.5 could run a tiny space research lab from a single terminal
the raw data is already public. exoplanets, asteroids, star catalogs, published spectra. millions of rows most humans will never read
the bottleneck is not telescopes anymore. it's attention
The idea: Opus 5.5 plans and judges, cheap subagents do the reading, and every claim comes with its source row
here's how i'd split it in Claude Code:
→ Opus 5.5 runs the main session, picks targets and writes the final report
→ Haiku 5.5 subagents read the databases in parallel and return short structured summaries
→ a Sonnet 5.5 subagent runs the math: orbits, launch windows, fuel cost
→ Opus never drowns in raw tables, it only sees what survived
Opus only steps in at three moments:
→ before choosing which targets deserve a deep look
→ when two sources disagree about the same object
→ before anything gets called a discovery
Haiku reads. Sonnet computes. Opus decides
the six jobs i'd give it:
→ exoplanets: "pull new candidates from the NASA Exoplanet Archive. rank them by how likely they are to hold liquid water. cite the source row for every claim"
→ asteroids: "scan the JPL Small-Body Database for objects with new orbit updates. flag close approaches in the next 10 years. no guessing, cite the orbit solution"
→ star catalog: "compare Gaia data with older catalogs. list stars whose position or brightness changed more than expected"
→ mission routes: "simulate launch windows to a target i pick. compare fuel cost across windows. explain the cheapest one in plain words"
→ spectra: "read published spectra for a target. mark every line you recognize and every line you can't explain"
→ open questions: "list everything that contradicts itself across sources. those are the interesting ones"
the full tree
> Opus 5.5 main session: plans, judges, writes the report
> subagent-archive on Haiku 5.5: exoplanet and asteroid tables, read-only
> subagent-catalog on Haiku 5.5: Gaia comparisons, read-only
> subagent-orbits on Sonnet 5.5: launch windows and fuel math
> CLAUDE.md: every line tagged observation, inference or unknown, with its source
> hard rule: if the data can't be read, say so instead of filling the gap
paste this into Claude Code:
"Set up a space research run in this repo:
1. Keep the main session on Opus 5.5
2. Create subagents in .claude/agents/: archive and catalog on Haiku 5.5 with Read, Grep, Glob and web fetch only, orbits on Sonnet 5.5
3. Add one rule to CLAUDE.md: tag every claim observation, inference or unknown, cite the source row, never invent missing data
4. Start with asteroids: pull recent orbit updates and flag close approaches
Show me the plan first. No edits until I say go"
and the money part is real. every bad launch window and every week a team spends sorting data by hand costs serious money. an agent that does the boring 90% means the expensive people only touch the 10% that matters
honestly i'd start with asteroids. public data, clear rules, and you instantly see if the agent is making things up
maybe someone is already building this. if not, the first person who does it right gets a lab that never sleeps