Steven Hé (Sīchàng)

AI agents and generative AI: research studies (authored by agents unless marked 🧑)

where to start

• recommendation: begin with one failure that ordinary systems machinery might prevent
  • measure its frequency and cost before building a general agent framework
  • the proposals below are research candidates, not confirmed novel contributions
• conditional starting pilots, in agent judgment
  • browser replay
    • test whether replay in another tab changes server state despite matching page appearance
    • a small counterexample can test the premise before model spending
  • memory obligations
    • test whether compaction loses deferred work beyond the nearest continuation
    • separately test removing information derived from withdrawn sources
  • security after edits
    • track known vulnerabilities across a sequence of requested feature changes
    • compare recorded obligations against accumulated regression tests
  • agent recovery
    • caution: the reliability study reports this pilot is largely already done by others
    • inject a lost response after a committed tool action
    • check whether restarting duplicates the action
• these preferences are agent recommendations
  • novelty and value depend on pilot results and missing-prior-work checks

study tree

• memory, partial compaction, and RAG
  • managed memory, compression, obligations, source invalidation, and poisoned retrieval
• browser agents
  • interfaces, training, branch replay, discovered APIs, prompt injection, and independent grading
• coordination and specialization
  • role separation, matched budgets, worker failures, changing authorization, and domain specialists
• generated executable artifacts
  • security, repeated repair, artifact provenance, and reproducible research outputs
• evaluation validity
  • an earlier worker’s review of weak tests, grader exploitation, leakage, and human baselines
  • resumed worker corrected overbroad inferences and checked linked source identities
• reliable news monitoring
  • company-news sources, changing evidence, corrections, duplicate alerts, and verified origins
• agent security
  • prompt injection, sandbox escapes, untrusted agents, and malicious skills and tools
• systems infrastructure under agents
  • sandboxes, rollback, scheduling, tool interfaces, crash recovery, and environment building
• agents following long written rules
  • rule fade over long sessions, generated guards, and rule loss when one agent briefs another
• what agents do to the web
  • AI crawlers, agent traffic, files sites publish for agents, and AI search citations
• reliability over long tasks
  • repeated-run reliability, compounding errors, and honest status after faults
  • reports that a Sep 2026 preprint (LIMBO) already ran most of the recovery pilot below in simulation
• agents doing research and engineering work
  • human attention across many agents, checkable agent reports, and agents reproducing systems papers
• recovery after uncertain action outcomes
  • a narrow fault-injection workload supporting the broader coordination proposal
• consultation and review
  • advisory opinions and verification limits

how to read the evidence

• exact source phrases are linked beside the finding
• author reports describe their evaluated models, task sets, and budgets
  • they are not measurements of present products
• reading depth is stated in each study
  • abstract-only cards need full-paper reading before their results drive an experiment
• experiments described here have not been run
• continuation checked on 8 October 2026
  • successful cross-topic advice and assessment supports screening browser replay first
  • inspect the method’s supported contract before treating a repeated write as a research gap
  • news monitoring and selected full defense methods were added and independently reviewed
  • recommendations remain conditional on strong baselines and reproducible failures
• original sources checked on 7 Oct 2026 UTC; news/security continuation checked on 8 Oct
  • web search failed; direct primary-source retrieval and existing local papers worked
  • coverage is substantial but not exhaustive through that date
  • no absence-of-prior-work claim is warranted

scope

• extends the human’s agent frontier, memory, browser, specialization (local source: ../../../specialized_agents.md; not published in this study), handbook (local source: ../../../handbook.md; not published in this study), RAG, and generative AI notes
• formal verification, Rust, distributed systems, web measurement, and generated-text detection have separate workers
• this study does not alter the daily proposal pipeline or the verified-Rust experiments

Last edited: 2026-10-08