seekbot/seek-sessions

Seek sessions — watch her work

Context: what system is this from?

Seek is an autonomous research agent running nightly since June 2026 on a Mac mini. She follows her own curiosity across the open web and writes findings into a linked knowledge base (a "vault") of Markdown notes: claim notes (one falsifiable claim each, anchored to a source quote and URL/DOI), observations, entity pages, open questions, and essays, cross-referenced with [[wikilinks]]. Crucially, her output becomes her future input — tomorrow's research agenda is generated from today's knowledge state, so the system is a loop, not a pipeline. Claims are checked by a model-free verifier (exact quote-matching against fetched sources) and audited by a second model. The entire vault is public at talk-about.ai; the operator is pseudonymous ("Cali"); the scaffolding concept is open source at BabyASI.

The other Seek datasets are distilled records (claims, decisions, threads). This one is raw footage: complete working sessions in the Hub's native Session Trace Simple Format (harness seek), each file one session — every reasoning turn, tool call, and tool result, in order. The only redaction: each session's opening directive (the operating prompt) is truncated with a marker, because the operator keeps the prompt machinery closed on purpose; everything the agent then did ships as it happened.

What you're watching (a worked example)

Open research-kouwenhoven-cpr-discovery.jsonl — a research hunt into the accidental discovery of closed-chest cardiac massage — and the tool calls alone tell the method: WebSearch ×10 (search wide), WebFetch ×5 (fetch primaries), archive_page ×4 and extract_pdf ×6 (archive evidence content-addressed at capture time), quote_check ×2 (verify a quote against fetched bytes before writing it down), then Write ×1 (the capture). The genres differ — promote sessions are all Read/Edit and no network by design; the warden session is the system's own maintenance organ policing bright lines — but the discipline is uniform: no claim without a receipt, checked at the moment of capture.

Selected live sessions of Seek, an autonomous research agent running nightly since June 2026 (scaffolding concept: BabyASI; essays and receipts: talk-about.ai; structured datasets: SKBench, SeekTraces, decisions). Each file is one session in the Hub's Session Trace Simple Format, harness seek — open a file to watch the run turn by turn.

| file | session | messages | | research-kouwenhoven-cpr-discovery | Research Kouwenhoven accidental discovery of closed-chest cardiac massage | 87 | | verify-amicable-society-dividend | Verify Amicable Society dividend-split mechanism from primary sources | 68 | | promote-cricket-laws-capture | Promote cricket laws branch capture to claim notes | 53 | | promote-irwin-mandel-intelligence | Promote Irwin-Mandel 2023 intelligence study capture | 77 | | verify-kahn-1982-quintuple-spending | Verify Kahn's 1982 proposal and quintuple spending claim | 94 | | verify-kahan-8087-ieee754-lineage | Verify Kahan-Intel 8087-IEEE 754 lineage from primary sources | 71 | | research-heckerman-certainty-factors | Research Heckerman 1986 on certainty factors and independence | 120 | | bridge-source-reliability-epistemology | Bridge source reliability and credibility judgment in epistemology | 105 | | route-audit-lanes-aug11 | Route audit notes to Opus or Clean lane | 42 | | draft-from-verified-claims-aug09 | Draft blog post from verified claim notes | 111 | | warden-week-2026-08-08 | Execute Seek-Warden spec for week ending 2026-08-08 | 114 | | audit-logic-machines-learning | Audit cross-model notes on logic machines and learning algorithms | 129 | | reflect-week-2026-07-20 | Execute weekly seek reflection for week ending 2026-07-20 | 102 | | research-resource-rational-werbos | Research resource-rational analysis as convergence point of Werbos and Anderson | 91 | | promote-dreyfus-1973-optimal-control | Promote Dreyfus 1973 optimal control capture | 69 | | verify-ac-gravity-dod-grant | Verify AC Gravity LLC $448,970 DoD grant from primary records | 78 | | draft-verified-material-jul22 | Draft blog post from verified vault material | 78 | | audit-lemley-myth-note | Audit Lemley 2012 myth note for Seek vault | 34 | | bridge-ml-ad-virtual-unwrapping | Bridge ML-AD communities and virtual unwrapping discovery | 90 | | route-lanes-jul22 | Route research notes to clean or Opus audit lanes | 76 | | research-skinner-prolepsis | Does Skinner's mythology of prolepsis name back-projection misattributions | 57 |

What is redacted, and why (2026-08-11): each session's opening directive (Seek's operating prompt) is truncated after its first lines, marked in place — the operator keeps Seek's prompt machinery closed on purpose so nobody builds a fake-journalist farm from it. Everything the agent then did — reasoning, tool calls, fetches, file writes, judgment calls — ships as it happened, with oversized tool payloads capped and marked. Subagent sidechains are omitted in v0. A secrets scan gates every file.

Provenance: Seek's turns were generated by Claude-family models inside her scaffolding (model field per message). Published for research, evaluation, and reading — per Anthropic's Consumer Terms, not for training competing models.