AI Engineer World's Fair 2026
Event notes (not attended — remote scan of the schedule + Online Track)

• When: June 29 – July 2, 2026 (Day 1 workshops → Days 2–4 sessions)
• Where: San Francisco, CA
• Organizer: AI Engineer (swyx & team)
• Attended: no. This note is a structured remote scan of the schedule + Online Track videos, filtered against corpus research themes.

Three canonical + one raw data snapshot; the three canonical URLs are the linkable references, the sources/ files are the exact bytes I filtered.

• worldsfair/schedule.pdf — official schedule PDF
• YouTube playlist — AI Engineer World's Fair Online Track 2026 (82 videos; only the "Online Track" subset is public)
• worldsfair/2026#schedule — official schedule page
• sources/ (this dir):
  • sessions.json (548 KB, schedule v4927, 556 sessions with {title, day, time, room, track, type, status, speakers, description})
  • speakers.json (977 KB, 630 speakers with bios)
  • session-details.json (672 KB, per-session descriptions + speaker bios indexed by numeric id 0..560)
  • youtube-playlist.json (89 KB, yt-dlp --flat-playlist --dump-single-json output for the 82 Online Track videos)
  • gen-tables.jq — reproducibly emits the org tables from sessions.json (jq -rf gen-tables.jq sessions.json > session-tables.org)
  • session-tables.org — 579-line generated table of all 556 sessions grouped by day (time · track · title · speakers · type)

The Fair is four days, ~556 confirmed sessions across ~15 named tracks

• workshop / sponsor / expo slots, and ~630 speakers. Day 1 is

workshop-heavy (58 sessions); Days 2–4 run ~166 sessions each. The Fair has dedicated tracks for three of the corpus's active research threads:

• Harness Engineering (13 sessions on Days 2–4) — the closest thing to a first-class venue for "harness = everything except the model" as an operational discipline. Overlaps directly with Aune's natea/harness-eval methodology and the CPRR frame.
• Sandbox & Platform Engineering (Day 3, 11+ sessions) — isolation primitives, agent runtimes, blast-radius containment. Overlaps with the sandbox + containment corpus and Kumar's secret-brokering thread.
• Memory & Continual Learning (Day 3) + Evals (Days 2–3) — memory as first-class harness surface; eval as the attribution instrument. Overlaps with the memory-systems + telemetry corpus and Poliakov's DBOS durability angle.

The Online Track (public YouTube) is 82 of the 556 sessions. Harness engineering dominates the online cut; memory + observability are present but thinner; secret custody, permission guardrails, and team topologies are essentially absent from the online set.

Three clusters, each anchored to a research thread in the wal.sh corpus. Sessions were filtered by four parallel agents against the sources/session-tables.org dataset (556 sessions). Overlap between clusters is noted inline (e.g. "Total Recall: Agent Memory and Harness Engineering" hits both memory and harness).

Each cluster is a curated list (10–30 sessions per cluster from the 556-session dataset), not the full track. The full per-day table lives at sources/session-tables.org if a broader scan is wanted.

82 videos out of the 556 sessions made the public "Online Track" playlist. This subset skews heavily toward harness engineering, eval, and observability — the sandbox / MCP / permission cluster is essentially absent from the online cut.

Curated below: the 17 videos with the closest bridge to corpus research. Full playlist metadata in sources/youtube-playlist.json.

• Poliakov (AI Tinkerers Boston, 2026-06-29) ↔ WF — Kay Malcolm (Day 4) "No Memory, No Harness: Why the Database Is the Last Line of Defense" is the closest DBOS-adjacent substrate talk. Viren Baraiya (Day 3) "Harnessing Agents: The Durable Runtime" complements it from the runtime side.
• Kumar (AI Tinkerers Boston, 2026-06-29) ↔ WF — Steve Yegge (Day 2) "Agentic Security: Permissions, Provenance, and the Agent Supply Chain" and Jim Clark (Day 2) "Who Approved That MCP Server?" are the governance analogs. Kumar's specific OPA-in-front-of-Secrets-Manager pattern is not directly matched.
• Aune (AI Tinkerers Boston, 2026-06-29) ↔ WF — Varun Krovvidi's "6 Pillars of an Agentic Harness" (Day 2), Rustem Feyzkhanov's "From Agent Traces to Agent Simulations" (Day 3), and Mike Krieger's "How Anthropic Builds" (Day 4) span the harness-methodology axis Aune measured empirically.
• Zhu (AI Tinkerers Boston, 2026-06-29) ↔ WF — the Fair is thin on Zhu's specific "deterministic code + structured memory" GTM-agent pattern; Nishant Gupta's "Deterministic Infra for Non-Deterministic AI Agents" (Online Track, 7m) is the closest at the substrate level.
• Team Topologies for the Agentic Platform ↔ WF — the Fair's track structure (Harness / Sandbox / Memory / Evals as first-class venues) is the four-hats model made external: the corpus argument that "boundaries survive teams" gets a concrete reflection in the Fair organizing its tracks around the same boundaries.

• [ ] Watch the ▶-linked Online Track picks in the table above; capture per-video notes as includes/<slug>.org for the ones worth verbatim citation.
• [ ] Add sources/schedule.pdf ingest if the PDF gains info not in sessions.json (e.g. room-map, ticket-tier gating).
• [ ] Cross-check the "Fair has 150 Anthropic-employees-in-title sessions" claim from Twitter buzz against ~sources/speakers.json (grep for Anthropic in the affiliation field, count).
• [ ] If natea/harness-eval methodology gets a WF write-up (Aune wasn't a WF speaker per the source data), link from the Boston Aune notes.
• [ ] Wire this note into site/events/index.org Upcoming/Past list.

Note assembled 2026-07-02 from a fixed dataset (sources/ files captured at scheduleVersion=4927 on 2026-07-02). Relevance filtering performed by four parallel scan agents against session-tables.org; the curation is this note's editorial judgment, not the Fair's.

Not attended; every session listed is a scheduled talk, not a witnessed one. Verdicts are attributed (sourced from the schedule payload) until a talk is actually watched from the Online Track, at which point the corresponding include should carry a :VERDICT: reproduced drawer.