Videos 4kYl2_mqmnQ
I Run a Fleet of AI Agents Across Three Machines. Here's What Broke. - Kyle Jaejun Lee, KRAFTON
Scene timeline
27 shot(s).
keyframes kept every frame deduplicated
What was stored
- cues
- 155
- whisperx 155
- chunks
- 17
- from 155 cues
- keyframes
- 23
- kept of 27 captured
- frames with text
- 23
- 752 lines read
- chapters
- 0
- from the source metadata
- keyframe bytes
- 2.0 MB
- word timings on 155 cues
Provenance
| stage | state | model | started | took |
|---|---|---|---|---|
fetch |
done | — | 2026-08-11 04:38 | 1m 09s |
stt |
done | — | 2026-08-11 04:40 | 10s |
chunk |
done | — | 2026-08-11 04:40 | 0s |
text_embed |
done | — | 2026-08-11 04:40 | 0s |
keyframe |
done | — | 2026-08-11 04:40 | 28s |
ocr |
done | — | 2026-08-11 04:40 | 13s |
frame_embed |
done | — | 2026-08-11 04:40 | 4s |
Frames, and what the machine read
-
- I run a fleet of Al agents across1.00
- three machines.0.99
- Here's what broke.1.00
- AI Engineer World's Fair 2026 - online talk0.99
- Kyle Jaejun Lee0.99
-
- The fleet1.00
- control plane1.00
- MacBook1.00
- Linux A0.97
- Linux B1.00
- heavy coding + personal projects1.00
- long-running coding tasks1.00
- short-lived personal side projects1.00
- macoS·sleeps0.94
- headless·always-on1.00
- headless·always-on1.00
- 3 machines0.99
- 2 Oses0.85
- sleeps vs always-on1.00
-
- The fleet, on screen1.00
- Overlord Gateway1.00
- Online1.00
- Dashboard1.00
- Map1.00
- Reviews1.00
- 18 active0.96
- 4 idle1.00
- 0 blocked1.00
- Context: avg -% |0.96
- Activity: 0 calls/5min0.99
- Context Heatmap1.00
- 40.77
- SYNCED1.00
- 18 active0.98
- 1idle1.00
- 4 blocked1.00
- 32 teams1.00
- CONTROL TOWER (4)0.98
- + ovl-ceo: 9 windows (attached)0.98
- L> + θ: zsh*: "Jaejunui-MacBookPro.local"0.96
- 0: 1 windows0.98
- tmux0.99
- で#20.51
- INFRASTRUCTURE (7)0.94
- bypass permissions on (shift+tab to...0.97
- 2 alerts0.97
- 0 panes0.99
- 0 panes0.96
- 0 panes0.99
- (4)1.00
- (M-a)1.00
- (M-b)1.00
- (M-c)0.95
- (5)0.92
- ovl-infra: 8 windows0.99
- convo-server: 1 windows0.98
- 1: vp-sysimprove!*:0.99
- 5:0.99
- 6:0.98
- 4:0.99
- : orchestrator-dev: "* Claude Code"0.95
- overlord-dev: "* Claude Code"0.94
- apple-note-mcp: "* Claude Code"0.97
- browser-ops: " Claude Code"0.95
- ccs-dev: "Jaejunui-MacBookPro.local"0.99
- security: "+ Claude Code"0.98
- "Jaejunui-MacBookPro.local"0.98
- "Inbox0.94
- ** bypass permissions on (shift+tab to c...0.97
- 0 panes0.99
- * bypass permissions on (shift+tab to c...0.98
- 0 panes1.00
- 0 panes0.99
- Enter to confirm0.99
- (M-k)0.99
- (M-d)0.99
- (M-e)1.00
- (M-g)0.99
- (M-h)0.98
- (M-1)1.00
- (M-j)0.99
- (M-1)0.96
- (M-m)1.00
- (M-n)0.97
- (M-0)0.98
- (M-p)1.00
- (M-q)0.99
- (M-r)1.00
- (M-s)0.88
- (M-f)1.00
- ovl-staff: 3 windows0.98
- ovl-personal: 18 windows0.98
- + 2: pmo: "* Claude Code"0.96
- + 0: node*: "Jaejunui-MacBookPro.local"0.97
- θ: zsh*: "Jaejunui-MacBookPro.local"0.94
- 1: rmo: "Jaejunui-MacBookPro.local"0.99
- 5:1.00
- 9:0.99
- 2:0.99
- 8:0.96
- 1:0.98
- 6:0.99
- life-dashboard: "* Claude Code"0.99
- zsh*: "Jaejunui-MacBookPro.local"0.98
- stacy-shift: "* Claude Code"0.94
- fabricator: "* Claude Code"0.96
- contents: " Claude Code"0.98
- re-agent: "+ Claude Code"0.95
- baby-care: "* Claude Code"0.95
- stock-textbook: "Jaejunui-MacBookPro.local"0.99
- tutorial: "* Claude Code"0.95
- happy: "Jaejunui-MacBookPro.local"0.98
- (M-t) - ovl-work: 1 windows0.97
- (sort:0.98
- (base)1.00
- base1.00
- : index)0.96
- jaejun.lee0.97
-
- The first thing that broke wasn't a0.99
- machine. It was me.1.00
-
- The overload1.00
- ctx 010.97
- ctx 020.99
- ctx 030.99
- me1.00
- scheduler1.00
- memory1.00
- reviewer1.00
- ctx 041.00
- ctx 051.00
- ctx 060.99
- 1 human · 3 roles0.96
- 4-6 live contexts0.98
-
- pane 4 (vp-aiewf-talk): EN0.96
- four standards locked, Phase 3-4 code (R14/R17/R18) committed, live migration dry-run resolved0.99
- pane 5 (home/2026-planning): EN0.99
- 0 ambiguous.0.99
- 4. Current gate (waiting): R18 live --apply is one-way-door → awaiting explicit CE0 go; next is0.98
- Discord吴星ユ.)0.89
- 6-pane §1 早(ovl-ceo:2) 0 至 → 歪.(pane 0 =,pane 2 =0.83
- 5. Open items: (a) confirm archiving 648 stale-tmp + 4 legacy dirs, (b) PR merge order0.99
- Phase 7 verifier → --apply (Mac) → Phase 5 fleet (krafton/personal) → Phase 8.0.98
- R10-R12/R13, (c) recommended separate ticket for the verifier.ts inb-${Date.now()} test flake.0.99
- 一,pane, ovl-personal0.84
- (Summary only - no files modified, no new work started.)0.99
- * Cogitated for 1m 16s0.99
- Brewed for 12s0.97
- CEO0.99
- Phase 7 verifier0.94
- Opus 4.8 ctx:1000k 30%used aiewf-talk [main]1.00
- Opus 4.8 (1M context)0.98
- ctx:1000k 50%used overlord [r3-spawn-verification]1.00
- bypass permissions on (shift+tab to cycle)0.99
- for agents1.00
- bypass permissions on (shift+tab to cycle)0.98
- . - for agents0.96
- [2026-06-25 11:48:15]1.00
- [2026-06-250.99
- [2026-06-250.99
- [2026-06-250.98
- [2026-06-250.98
- [2026-06-251.00
- 11:48:18]1.00
- 11:48:24]1.00
- 11:48:30]0.99
- 11:48:21]0.97
- 11:48:27]1.00
- Error1.00
- Error1.00
- Error1.00
- Error1.00
- Error1.00
- Error1.00
- polling channel1.00
- polling1.00
- polling0.97
- polling0.97
- polling1.00
- polling1.00
- channel1.00
- channel1.00
- channel1.00
- channel1.00
- channel1.00
- 1484342634523660319:0.99
- 1484342634523660319:1.00
- 1484342634523660319:0.99
- 1484342634523660319:1.00
- 1484342634523660319:1.00
- 1484342634523660319:1.00
- getaddrinfo ENOTFOUND discord.com1.00
- getaddrinfo1.00
- getaddrinfo1.00
- getaddrinfo1.00
- getaddrinfo1.00
- getaddrinfo1.00
- ENOTFOUND1.00
- ENOTFOUND1.00
- ENOTFOUND1.00
- ENOTFOUND1.00
- ENOTFOUND1.00
- discord.com1.00
- discord.com1.00
- discord.com1.00
- discord.com1.00
- discord.com1.00
- ●Here's the 5-line summary based on vp-llm-gateway-aws's commander-intent.md and status.md:0.99
- 2. Current progress (GREEN): PR #100 is complete - v2 family-aware routing0.98
- 1. Mission: Set up the collaboration structure for Krafton's AwS Bedrock-based LLM Gateway PoC,0.99
- delegate implementation to workers, and sync with human collaborators (Kwangyoon on Gw admin API,1.00
- Woojin on GetAPI calls)0.98
- - running in parallel with the CE0's manual Bifrost test track.0.98
- [2026-06-250.98
- 11:48:33]1.00
- Error1.00
- polling1.00
- channel1.00
- 1484342634523660319:1.00
- getaddrinfo1.00
- ENOTFOUND1.00
- discord.com1.00
- (tool/vision/field-matching) on /endpoint-v2, meeting both CE0 requirements (v1 left untouched +0.99
- [2026-06-251.00
- 11:48:36]1.00
- Error1.00
- polling0.98
- channel1.00
- 1484342634523660319:1.00
- getaddrinfo1.00
- ENOTFOUND0.99
- discord.com1.00
- OpenAI/Anthropic field matching); now awaiting CEO review.0.98
- [2026-06-250.98
- [2026-06-251.00
- 11:48:39]1.00
- 19:40:58]1.00
- Error1.00
- Error1.00
- polling1.00
- polling1.00
- channel1.00
- channel1.00
- 1484342634523660319:0.99
- 1484342634523660319:1.00
- getaddrinfo ENOTFOUND discord.com1.00
- Failed to parse1.00
- 3. **Verification1.00
- Interrupted · What should Claude do instead?0.98
- comander-intent.mdstatus,md,0.93
- 答,号0.54
- [2026-06-251.00
- 21:22:25]1.00
- Error1.00
- polling1.00
- channel1.00
- 1484342634523660319:0.99
- Failed to parse0.99
- updates1.00
- Opus 4.8 (1M context)0.97
- ctx:1000k 5%used llm-gateway-aws [main]1.00
- bypass permissions on (shift+tab to cycle)1.00
- for agents1.00
- 3. Done0.94
- - v10 five sections + main title LOCKED, capture-brief-vll written, two real receipts0.98
- confirmed within budget1.00
- _+ overall TOP 7 selected + public-transit commute times measured.0.97
- secured (F1 real-mac-panic.png, F4 real-mem-pressure.png, both scrubbed/verified), plus 6 existing1.00
- 3. Phase 3-1 done: TOP 7 commute grades finalized (Seongdong Samsung Chèrevill = top A-grade),0.99
- HAVE captures + 2 new frames.1.00
- redevelopment surveyed (only Jeonnong Woosung promising), Threads 43 filtered (10 new A-grade0.99
- completion is its prerequisite), and overlord worker spawn launches gpt-oss-120b instead of1.00
- Claude, blocking task delegation - VP must drive the pipeline directly for now.0.98
- 5. Next steps - ① CE0 produces 5 design diagrams + F2/F3 mockups; ② apply 50-§2 voice comments,0.97
- 4. Blockers - The human-voice VO recording is gated on CEo's time (the biggest variable; script0.97
- for the 10 A-grade Threads finds (Gwanak Bongcheon first) ③ verify price/area of 9 R0.98
- 4. Next steps (backlog): ① jeonse-ratio/gap-investment analysis (apt2.me) ② commute measurement0.98
- A-promotion ④ design field-visit routes ⑤ re-verify northwest region on Hogangr0.98
- found).1.00
- for1.00
- then finish the §3·§4-§5·close script (correct sleep/00M → "kernel panic"); ③ assemble the v100.96
- 5. Key rules: split independent tasks into separate tmux panes, no CDP (ager0.99
- video → record human-voice VO + sync + CEO approval → submit.0.96
- only), every number must be backed by a screenshot.1.00
- * Cogitated for 9s0.99
- * Worked for 9s0.98
- Show me the v10 script draft so far0.99
- start the jeonse-ratio analysis with apt2.me0.99
- Opus 4.8 (1M context) ctx:1000k 5%used vp-aiewf-talk [main]0.99
- Opus 4.8 (1M context) ctx:1000k 4%used 2026-planning [main]0.99
- bypass permissions on (shift+tab to cycle).- for agents0.98
- bypass permissions on (shift+tab to cycle)0.99
- for agents0.93
- [ovl-ceo] 2:aiewf-talk* 4:[tmux]- 5:vp-llm-gateway-aws0.98
- 7:vp-ludo-quali0.99
- the mess that broke me1.00
-
- The insight1.00
- a few execs run a1.00
- company of thousands1.00
- separate context1.00
- hierarchy1.00
-
- State lives in files0.99
- ~/.agents/overlord/1.00
- shared/1.00
- #0.99
- everyone must honor1.00
- workspaces/{entity}/1.00
- # each agent's desk0.99
- commander-intent.md· status.md· handoff/0.98
- machines/{host}/{team}/ # machine-bound state0.99
-
- Inbox1.00
- Pending Reviews81.00
- PENDINGvp-sysimprove1.00
- R15:Paperclip 一, read-only mirror GO0.93
- 2 Review: R6: Durable Dispatch Queue + boot drain0.97
- xp-syumgrove0.76
- requested 2d ago0.98
- Taquoid by vp-eysmprecmanaget0.59
- Tr v-eympreve 4d ago Tuck 50u55c30.55
- Chongos fram prevloen review:0.91
- Toggle don0.79
- PENDING1.00
- vp-sysimprove1.00
- R6: Durable Dispatch Queue + boot drain1.00
- requested 4d ago1.00
- TL;DR (CEO 1 )0.96
- PENDING vp-sysimprove0.98
- queue 00 tasks durable queuepending→dispatched→acked+_progressverifying')0.74
- ACKpolAckI'isAcked'),rconclerack-tmeoutrenotry0.72
- ackedfase)(tasks,t:46,470,475°).tasdspatched , 20.60
- R7: Registry↔Reality Reconcile Loop + cron1.00
- reconcleracktimeout enoty交网Reunchd 对0.74
- X task drain 早色dispatchedask早世号至百.3.renaotiy *0.75
- requested 4d ago0.99
- Acd ine comemest.0.84
- PENDINGvp-issue-tracker1.00
- Coros0.66
- interlock)1.00
- issue-tracker v0.1 — research + plan (GO/ports/workers/R1-0.98
- [6/20]*:R3/R4/R5dispetch30[vp-sysimprove)单dispatched→空0.85
- → CEO 中层对AL**型期响 G1+G2+03 每 繁*0.63
- requested 8d ago0.98
- *+**: queue型型((durable).**[a)dispatchpending-noty,()spawn率0.79
- ****:= **1.5-2%, -150 LOC**, 2 0, 单 two·woy-doo: R7 reconcle/ 弯斗三.0.72
- R1: Cross-machine Worker (Hybrid-lite, Panarchy )0.99
- PENDING vp-sysimprove0.98
- Dart 1_ dinnntnh ol mnn / s0.65
- Approve0.90
- requested 8d ago1.00
- PENDING1.00
- vp-sysimprove1.00
- R2: Permission auto+escalation (gate-check 誓-考)0.96
- one approval cycle1.00
-
- Then one machine wasn't enough.0.99
-
- Failure 1 agents do the work instead of dispatching it0.99
- sh-3.2$ overlord dispatch --help0.99
- Usage: overlord dispatch [options] <team> <message-or-file>1.00
- Dispatch a task to a team (creates a task + trace, notifies manager)0.99
- Arguments:1.00
- team1.00
- Target team name1.00
- message-or-file1.00
- Task message or path to a .md file1.00
- orchestrator layer0.99
- does the work itself0.98
- G0.98
- Options:0.99
- --priority <level>1.00
- --from <sender>0.98
- Priority: low/normal/high/urgent (default: "normal")1.00
- Sender name (default: "ceo")1.00
- => wrong!1.00
- display help for command0.99
- CLI harness forces dispatch1.00
- CLI to control1.00
- Worker1.00
- /overlord-ceo1.00
- Overlord CEO session - organization management, team0.98
- dispatch, status monitoring. Auto-load when operating as.0.99
- /overlord-team-manager1.00
- Overlord team manager session - team operations, worker1.00
- management, status reporting, Auto-load when operating a0.98
- )/overlord0.95
- Skills to call CLIs0.99
-
- Failure 20.96
- panes too small1.00
- tmux (tmux)1.00
- 10.99
- ..ptures-manual (-zsh)0.99
- 220.86
- periodically and summarize1.00
- haven't used in a while. They0.97
- (#1, #2, #7, #9, #6) for you to0.97
- 411.00
- export default fun1.00
- results as they land?1.00
- still add startup and1.00
- review.1.00
- ction RootLayout({0.99
- * Cogitated for 4m 38s0.99
- with /plugin1.00
- context cost1.00
- - review them0.99
- * Sautéed for 2m 36s1.00
- Baked for 8m 1ls · 1 shell still0.96
- running1.00
- Contemplating- (11m 26s)0.95
- !0.83
- Draft a remediation plan0.98
- intent. Adds a small piece of1.00
- 2>/dev/null; echo "=====0.98
- any authenticated user with no1.00
- (ctrl+b ctrl+b (twice) to run1.00
- app code (guarded off in1.00
- prod).1.00
- Running 1 shell command..0.96
- dynamic =====" && gr-0.93
- special access.1.00
- in background)1.00
- Transfiguring_ (11m 26s)0.96
- Percolating.. (11m 25s)0.96
- Cogitated for 3m 59s1.00
- Garnishing... (1lm 24s)0.96
- 5. Chat about this1.00
- Enter to select · Tab/Arrow keys0.98
- to navigate · Esc to cancel0.98
- in background)1.00
- (ctrl+b ctrl+b1.00
- (twice) to run0.99
- es/baby-care-history/fix-invi1.00
- te/app/actions/setup.ts1.00
- isolation, untouched by other1.00
- stash), commit it in1.00
- * Verifying npm run build..0.97
- agents. Safest.1.00
- * Enlarging touch targets...0.99
- * Grooving... (1lm 20s)0.93
- Crafting... (1lm 17s)0.97
- 6. Chat about this0.98
- Enter to select1.00
- Esc to cancel1.00
- · t/i to navigate0.90
- 15 }0.96
- and I'll scope it with branches:1.00
- $ cat1.00
- agent is unaffected.1.00
- Allowed by auto mode0.99
- Running 1 shell command..0.99
- [main].1.00
- /tmp/ts-strict-errors.log1.00
- Lint/format fixes get a clean1.00
- PR against main.0.99
- * Julienning.. (11m 15s)0.97
- Churning.. (thought for 2s)0.98
- * Sautéed for 6m 38s1.00
- Bloviating.. (thought for 2s)0.97
- 5. Chat about this0.99
- push it and open the PR0.99
- Enter to select · Tab0.97
- navigate · Esc to0.96
- care-history/app/(main)/mo1.00
- 2>/dev/null; sleep 2; ps aux0.97
- lery-opt-verify && npm run0.98
- queue module1.00
- File must be read first0.99
- re/page.tsx)1.00
- build" | grep baby-care |-0.95
- | grep "next/dist/bin/next0.98
- (ctrl+b ctrl+b (twice) to run1.00
- in background)1.00
- build 2>&1 | tail -450.97
- just these 41.00
- testable, ma1.00
- * Noodling... (1lm 7s)0.94
- Baking- (11m 7s)0.96
- Dilly-dallying.. (1lm 4s)0.96
- * Verifying build... (1lm 2s)0.97
- records'.1.00
- 4. Chat about th0.99
- Enter to select1.00
- navinate . Fer0.95
-
- Failure 31.00
- out of memory0.99
- 2.1.1911.00
- 525.4MB1.00
- 361.00
- 911.00
- 158261.00
- ja0.98
- 2.1.1911.00
- 529.1MB1.00
- 361.00
- 1031.00
- 181241.00
- ja0.97
- 2. 2.1.1911.00
- 490.5MB1.00
- 361.00
- 1031.00
- 104021.00
- ja0.99
- 2.1.1911.00
- 304.6MB1.00
- 191.00
- 741.00
- 221291.00
- jat0.85
- 2.1.1911.00
- 397.0MB1.00
- 331.00
- 1001.00
- 104031.00
- ja0.98
- 2.1.1911.00
- 467.0MB1.00
- 331.00
- 1011.00
- 421741.00
- ja0.98
- 2.1.1910.99
- 476.8MB1.00
- 361.00
- 1031.00
- 466731.00
- ja0.99
- 2.1.1911.00
- 492.4MB1.00
- 221.00
- 771.00
- 607591.00
- ja0.99
- 2.1.1911.00
- npm exec @upstash/context7-mcp1.00
- 54.9MB1.00
- 121.00
- 501.00
- 778551.00
- jaejun.lee0.97
- 2.1.1911.00
- npm exec @upstash/context7-mcp1.00
- 54.9MB1.00
- 121.00
- 501.00
- 182161.00
- jaejun.lee1.00
- 2.1.1911.00
- npm exec @upstash/context7-mcp0.99
- 54.8MB1.00
- 121.00
- 501.00
- 779161.00
- jaejun.lee0.97
- 2.1.1911.00
- npm exec @upstash/context7-mcp1.00
- npm exec @upstash/context7-mcp1.00
- 54.7MB1.00
- 54.6MB1.00
- 121.00
- 121.00
- 501.00
- 501.00
- 382161.00
- 467001.00
- jaejun.lee1.00
- jaejun.lee1.00
- 2.1.1910.96
- npm exec @upstash/context7-mcp1.00
- 54.6MB1.00
- 121.00
- 501.00
- 133421.00
- jaejun.lee1.00
- npm exec @upstash/context7-mcp1.00
- 54.5MB1.00
- 121.00
- 501.00
- 978101.00
- jaejun.lee1.00
- Open and Save Panel Service(Claude)1.00
- 54.5MB1.00
- 71.00
- 3631.00
- 492761.00
- jaejun.lee1.00
- npm exec @upstash/context7-mcp1.00
- 54.4MB1.00
- 121.00
- 501.00
- 422001.00
- jaejun.lee1.00
- npm exec @upstash/context7-mcp1.00
- 54.3MB0.93
- 121.00
- 501.00
- 104901.00
- npm exec @upstash/context7-mcp1.00
- 54.3MB1.00
- 121.00
- 501.00
- 132891.00
- npm exec @upstash/context7-mcp1.00
The page's on-screen-text budget of 600
lines is spent, so the last cards in this grid list fewer lines than they
hold. Narrow the page with ?frames= to read them.
Transcript
155 cues· 1,653 words· 9,000 chars
- 0:00 Hello, I'm Kyle.
- 0:01 I run a fleet of AI coding agents across three machines every day, and this is the story of what broke, starting with me.
- 0:08 This is my fleet.
- 0:09 Three machines I run every single day.
- 0:12 The MacBook is where I do my heavy coding and my personal projects.
- 0:16 And it sleeps because, well, it's a laptop.
- 0:18 Then two Linux boxes, both headless and always on.
- 0:22 Linux A takes the long-running coding tasks.
- 0:25 Linux B runs my short-lived personal side projects.
- 0:28 Two operating systems, one control plane tying it all together.
- 0:32 And to be clear, this isn't a demo I spun up for the talk.
- 0:35 This is what I actually use all day, every day.
- 0:39 And here it is live, the whole thing running on one screen.
- 0:42 Real agents, real work.
- 0:43 This is a normal Tuesday for me.
- 0:45 So let's go back to the beginning, because before any machine broke, I broke.
- 0:49 When I started, it was just me and a terminal.
- 0:53 A few TMUX panes, a few agents, then a few more.
- 0:56 And really fast, I'm sitting in front of four, five, six live contexts at the same time.
- 1:02 And here's what nobody warns you about.
- 1:04 At that point, I'm not running agents anymore.
- 1:07 I've become the scheduler, deciding who does what.
- 1:09 I'm the memory.
- 1:10 holding what every one of them is doing, and I'm the reviewer, checking all of it.
- 1:14 One human, three roles, six contexts.
- 1:17 It does not scale.
- 1:19 This.
- 1:20 This is the mess that broke me.
- 1:22 I just couldn't hold what six agents were doing at once.
- 1:25 My own attention was the bottleneck.
- 1:28 So, I asked myself kind of a strange question.
- 1:31 How does a handful of executives run a company of thousands of people?
- 1:34 They don't hold all of it in their heads.
- 1:36 They separate context.
- 1:38 Each person only ever sees their own slice.
- 1:40 And that was the unlock.
- 1:41 What if my agents weren't a flat pile?
- 1:44 What if they were an organization?
- 1:46 So that's what I built.
- 1:47 A hierarchy, CEO, VP, manager, worker.
- 1:50 And please hear me on this.
- 1:51 These are real entity types in the system.
- 1:54 It is not a cute metaphor.
- 1:55 Each one is its own agent with its own scoped context and its own approval boundary.
- 2:00 Context flows down.
- 2:02 Each layer only gets the slice it needs.
- 2:04 Results flow back up, and I only review what reaches the very top.
- 2:08 So instead of holding six contexts in my head, I hold exactly one.
- 2:12 x problem where does an agent's state actually live normally it lives inside the model's context window and that window fills up so i moved it out every entity gets its own workspace on disk shared context that everyone has to honor lives in shared machine bound state lives under machines and inside each workspace the mission the current status and a handoff folder the actual work product that gets passed along the state lives in files it is not trapped inside one model
- 2:43 And this is the single most practical thing I learned all year.
loading