BootLoops, from Harvard physicist Matthew Schwartz, is an open-source harness of exact-calculation tools and agent skills for quantitative science. Its GitHub org has 7 public repos, mostly MIT-licensed, including a skills repo packaged as Claude Code and Codex plugins.
How we checked · Why it matters · Try this
How we checked
bootloops.ai and GitHub org list open-source repos and licenses; code not run
Why it matters
It is a public example of a domain-specific harness and skill set built for any model doing long, checkable research work.
Try this
Browse the BootLoops-ai skills repo on GitHub and read its research protocols as a model for your own agent skills.
T3 Code merged PR #28410 on Nov 2, 2026. Claude processes it starts now report CLAUDE_CODE_ENTRYPOINT as t3-code instead of the Agent SDK default sdk-ts.
How we checked · Why it matters · Try this
How we checked
PR page shows merged Nov 2, 2026 with entrypoint change; not run
Why it matters
Apps are relabeling their entrypoint to avoid how usage may be weighted. The linked issue warns this hides the label and may conflict with Anthropic's terms.
Manus 2.0, announced Sept 28, 2026, adds the Cascade agent harness, Automations that start from events such as a new email or calendar event, and an always-on Cloud Computer. Video Editor followed on Oct 1.
How we checked · Why it matters · Try this
How we checked
Manus blog confirms 2.0 features and dates; Cascade parallel sub-agents not confirmed
Why it matters
Event-triggered agents that run on persistent hosted machines are now a packaged feature, not something you have to build yourself.
Try this
Read the Manus 2.0 post to see which connected services Automations can trigger on before designing a workflow.
Anthropic says protein design and other dual-use biology research features are not available for general access in Claude Fable 5. It plans an access program for scientists, and Opus 5 remains its most capable generally available model.
How we checked · Why it matters
How we checked
primary link exists
Why it matters
Biology builders should expect refusals or blocks on the newest model and plan around Opus 5 or a future trusted access program.
To run many long Claude Code projects at once, Schwartz used one session per project and a master session that coordinates the others. Background subagents meant a classifier block stopped only one agent, and periodic file consolidation helped against context loss from compaction.
How we checked · Why it matters · Try this
How we checked
primary link exists
Why it matters
These are practical patterns for multi-day agent work: isolating failures, a coordinator session, and files as durable memory across compaction.
Try this
Add a step to long agent tasks that has the agent consolidate its notes into one up-to-date plan file at regular intervals.
In Vibe Physics (first published March 2026), Schwartz had Claude Code keep a tree of markdown files with one summary per stage and one file per task. He also cross-checked its work with GPT and Gemini and added a CLAUDE.md rule banning skipped steps.
How we checked · Why it matters · Try this
How we checked
primary link exists
Why it matters
These are simple, reusable ways to cut silent errors and lost context when an agent works through a long technical project.
Try this
Add to CLAUDE.md: never use phrases like 'this becomes' to skip steps; show the calculation or say 'I don't know'.
Anthropic reports that Mythos Preview and Opus 4.8 designed protein binders that worked against 14 of 15 targets in outside wet-lab tests. Hit rates were 22.6% to 35.1%, against a typical 10-15%.
How we checked · Why it matters · Try this
How we checked
Post exists and states these figures; wet-lab results are Anthropic's report, not independently checked
Why it matters
This is a concrete public test of an agent running a long, multi-model science pipeline on its own, and the prompts and data are released.
Try this
Read the released 30,000-token protein design prompt to see how Anthropic set up a 48-hour autonomous agent session.
Anthropic says Claude Opus 5 in Claude Science took raw proprietary NMR and LC-MS instrument files and a short prompt, with no vendor software. It returned results in 23 and 19 minutes, with purity 96.4% versus the lab's 96.33%.
How we checked · Why it matters · Try this
How we checked
Post exists and states timings and purity match; single-sample result reported by Anthropic
Why it matters
It suggests a generally available model can decode undocumented binary formats and do routine lab analysis, with self-checks against the file's own totals.
Try this
In Claude Science, give Claude a raw NMR or LC-MS file and ask it to confirm the compound's identity and purity.
Ray Fu published a free guide to building a multi-agent stock research team in Manus, with Analyst, Quant, Risk and Report agents. Its triggers include earnings dates, newsletters and CEO posts on X.
How we checked · Why it matters · Try this
How we checked
Guide exists; trading value not assessed; X-post trigger not confirmed in Manus docs
Why it matters
The guide exists, but its trading usefulness is unproven and its X-monitoring trigger raises platform terms concerns.