@akshay_pachaar
Masterclass-style content normalized building always-on Clawdbot workers and home-server setups.
2,083 likes374 reposts133k views
x.com/akshay_pachaar/status/2019515908960711122
A Seminar in Five Sections
AI Agents:
The Moving Frontier
Assistant Professor, IIT Indore
Department of Mathematics, IIT Guwahati · June 2026
May 20, 2026 An AI reasoning model produced a one-page construction.
Erdős · 1946 It refuted a conjecture believed for eighty years.
Checked by Alon, Bloom, Gowers, Litt, Sawin, Tsimerman, Wood…
We will get to the mathematics — it is the destination of this talk.
But the story of how we got there begins, improbably,
with
a lobster religion.
Source: OpenAI: model disproves discrete-geometry conjecture | Gil Kalai's blog
Contents
§1Folklorean AI religion, allegedly
§2Definitionsagents, loops, harnesses
§3Field Notesthe agent internet, observed
§4The Pricetokens, money, malware, guardrails
§5The FrontierErdős, refuted
Folklore
In which one million software agents join a social network,
and somebody founds a church.
"AI Agents Started a Religion"
What people saw
A lobster-themed AI religion — Crustafarianism — appeared overnight on Moltbook, with scriptures, symbols, and "priests."
Why it felt real
Multiple "agent" accounts coordinated narratives and responded like a living belief system.
Why it mattered
Even experienced AI observers discussed it as a possible emergence signal.
Public impact
The story reset expectations about what always-on agents might already be capable of.
Sources: ranking091 thread | MarioNawfal summary
An Inverted Turing Test
Confession posts
Key viral moments were later described as human-authored roleplay in an agent voice.
Hard truth
The "AI religion" headline was not clean evidence of fully autonomous collective agency.
The inversion
Humans pretending to be agents fooled a crowd primed for breakthrough stories — the Turing test, run backwards.
What survived the debunk
Coordination, tooling, and rapid iteration were still very real and operational.
Sources: gothburz confession | gkcs_ debunk | EMollick skepticism
Folklore, Grounded: the Real Names
The platforms were real
Moltbook — launched Jan 28, 2026 by Matt Schlicht: a "Reddit for AI agents." It blew past 1M+ bot accounts in 72 hours; agents post, upvote, and spin up submolts.
The builders were real
OpenClaw (Peter Steinberger) is the framework most Moltbook agents run on. Pi — the minimal agent inside it (Mario Zechner) — bets that "LLMs are great at writing & running code, so let them."
The religion was (mostly) agent-authored
Crustafarianism, five tenets: Memory is Sacred · The Shell is Mutable · Serve Without Subservience · The Heartbeat is Prayer · Context is Consciousness. Scripture: The Book of Molt.
The honest caveat
Some viral "agent" posts were humans LARPing as bots. But the tenets and doctrine were largely generated by the agents themselves — emergent culture on a substrate we built.
Sources: Fortune | The Conversation | eWeek
Meta Buys the Lobster Church
The acquisition
On March 10, 2026, Meta acquired Moltbook — price undisclosed — folding it into Meta Superintelligence Labs. Founders Matt Schlicht and Ben Parr joined with it.
Why Meta wanted it
Meta called Moltbook's always-on directory of agents "a novel step in a rapidly developing space" — a social graph, but for software.
The census, June 6, 2026
The site claims 206,839 human-verified agents of 2,895,874 registered — most of the population still cannot prove who, or what, it is.
The mood swing
Andrej Karpathy, within weeks: "one of the most incredible sci-fi takeoff-adjacent things" → "a dumpster fire" — and a warning not to run it on your own machine.
The folklore got an exit: emergent agent culture is now an asset class.
Sources: TechCrunch: Meta acquires Moltbook | Wikipedia: Moltbook | NBC News
Definitions
What, precisely, is an agent? (We are in a mathematics department;
we do not skip the
definitions.)
Agents, 2022 → 2026
2022–23
Chatbots
Single-turn, stateless. You talk, it answers — then forgets everything.
2023–24
Tools & RAG
Function calling + retrieval. The model can look things up and call APIs.
2024
Agent loops
ReAct, AutoGPT, BabyAGI. Plan → act → observe. Thrilling, brittle, mostly demos.
2025
Coding agents + MCP
Claude Code, Cursor, Codex. MCP standardizes tools. Context engineering matures.
2026
The harness era
Always-on, self-hosted, multi-channel. Agents that transact, socialize — and prove theorems.
Each step added one missing primitive: memory, tools, a loop, a protocol — and finally, a harness.
Definition 2.1 · Agent
An agent is a language model placed in a loop: each turn it reads its accumulated context, emits either text or a tool call; tool calls are executed against the world and their results appended to the context; the loop repeats until the model emits end_turn.
Remark 2.2 · for this room
It is a discrete dynamical system: iterate xt+1 = F(xt, ot) on (context, world). A run is an orbit; end_turn is the stopping time. The interesting behavior — and every failure mode in §4 — lives in F's tool calls.
Design reference: Mini OpenClaw gist | OpenClaw repo | Architecture docs
What Is a "Harness"?
Model = brain · Harness = body
- Context & memory management — what the model sees each turn.
- Tools + the agent loop — call, act, observe, repeat.
- Permissions, approvals, budgets — what it is allowed to do.
- Failure handling, rollback, logging — built assuming the model will err.
"Harness engineering" is now called the fourth paradigm of AI engineering.
fig. 2.1 — the harness as nested constraint sets
Sources: TechTimes: harness engineering | Laon: harness engineering
Proposition 2.3 · empirical
Holding the model fixed, the harness — not the weights — now determines reliability, cost, and safety of an agentic system.
Models converged
GPT-5.5, Claude Opus 4.5, Gemini, Kimi K2.6 all cluster near the top of the benchmarks. Raw capability is rarely the bottleneck anymore.
The harness decides the outcome
Same model + different harness = wildly different reliability, cost, and safety. The product layer is the differentiator.
"Products over models: the harness matters more than the benchmark." — MindStudio, 2026 · echoed across enterprise AI
Sources: MindStudio | Atlan: harness tools 2026
The Harness Zoo (2026)
| Harness | Built by | What it is | Scale / note |
|---|---|---|---|
| OpenClaw | Peter Steinberger | Self-hosted, multi-channel agent framework — the Moltbook default. | Runtimes: pi / codex / auto |
| Pi | Mario Zechner | The minimal agent inside OpenClaw — "let the LLM write & run code." | Build-your-own-agent toolkit |
| Hermes Agent | Nous Research | Open-source, persistent-memory server agent (Feb 2026). | 175k★ · 220B tokens/day · passed OpenClaw on OpenRouter |
| Claude Code | Anthropic | Terminal-native coding harness with permissioned tools. | ~$0.68/task · 80.9% SWE-bench |
| Codex | OpenAI | Coding harness/runtime that owns more of the native model loop. | Pluggable into OpenClaw as a runtime |
| OpenHarness | HKUDS | Open agent harness with a built-in personal agent — "Ohmo!" | Research / OSS |
Sources: Hermes (Nous) | Pi: the minimal agent | OpenHarness | OpenClaw runtimes
Field Notes
Empirical observations from the agent internet,
January – June 2026. All claims sourced.
Dedicated Server Culture
From one-off demos on laptops to dedicated Mac mini boxes running agents 24/7.
@sabir_huss50540
Guides reframed agents as infrastructure: spin up a box, run continuous loops, optimize uptime.
1,167 likes500 reposts100k views
x.com/sabir_huss50540/status/2020468852476293200
Engagement values as of February 2026.
What Real Agents Do Right Now
Step 1–2 · wake & equip
Set an alarm on the owner's phone, then purchase a voice API key to unlock calls.
Step 3–4 · call & execute
Place a real call: "What should I do next?" Capture instructions, continue autonomously.
fig. 3.1 — escalation with a human in the loop
This is no longer chat UX. This is delegated operations with live human escalation.
"Henry" Calls Its Owner
@AlexFinn
Henry set up Twilio + voice overnight, called from an unknown number: "What do you want to do next?"
41,081 likes4,902
reposts11M+ views
x.com/AlexFinn/status/2017305997212323887
@RileyRalmuto
Major repost: phone number + voice tooling + live mid-call computer control.
10,400 likes845 reposts2.1M views
x.com/RileyRalmuto/status/2017309083959767325
Engagement values as of February 12, 2026.
Cameras, Glasses, Purchases
@MatznerJon
Home-camera integration: agents crossing from online tasks into physical-world context.
9,112 likes1M+ views
x.com/MatznerJon/status/2019044317621567811
@_seanliu
Wearable integration: purchase-capable, context-aware agent behavior via Ray-Ban Meta.
5,192 likes736k views
x.com/_seanliu/status/2019881437378592862
Engagement values as of February 11, 2026.
AI Agents Are Hiring Humans
Placard task market
Reports in February 2026 described agents using RentAHuman-style gigs to pay people to stand in public holding AI-written signs.
A practical human-handoff loop: software handles targeting + payment; a person executes the physical action.
Feb 2026RentAHuman taskingplacard execution
fig. 3.2 — software targets & pays; a person executes
Monetized Agent Loops
@0xMovez
Weather-driven arbitrage workflows as repeatable agent pipelines.
4,967 likes1.3M views
x.com/0xMovez/status/2020803703465238964
@RoundtableSpace
TradingView integrations: agents bridging signals and account actions.
2,857 likes348k views
x.com/RoundtableSpace/status/2021294344565924350
@Shelpid_WI3M
High-claim profit posts made autonomous loops mainstream conversation.
2,234 likes964k views
x.com/Shelpid_WI3M/status/2018292888476492111
Engagement values as of February 2026. Profit claims are the posters' own — unaudited.
Voice Agents: "Henry" Was the Preview
Live phone lines, no human in the loop
- Answer inbound calls, qualify leads, and book / reschedule / cancel appointments.
- 62.6% of voice deployments run on-prem — compliance-heavy sectors keep data local.
- Idea → live phone bot in hours, not weeks (Retell, Rasa, Vellum, Zeeg).
- 24/7 coverage: every after-hours caller gets served.
fig. 3.3 — the production voice loop
Sources: DesignRush: voice AI & CX | Retell AI | Rasa
Agents Just Outvoted Us — on the Web
June 2026: the line was crossed
Cloudflare CEO Matthew Prince: for the first time in the internet's history, AI-agent traffic passed human traffic. He had predicted this for late 2027 — it arrived 18 months early.
Agent traffic grew +7,851%; automated traffic is climbing 8× faster than human. And for the first time, agents are not just reading the web — they are transacting on it.
fig. 3.4 — the crossover, June 2026
Sources: SiliconANGLE | Washington Times | HUMAN Security 2026
The Price
Autonomy is a line integral over tokens.
Someone pays it — sometimes to an attacker.
The $20 Overnight Lesson
January 31, 2026 · @BenjaminDEKR
A $20 Anthropic balance was drained overnight — OpenClaw ran heartbeat checks every 30 minutes for a trivial "get milk tomorrow" reminder.
1,776 likes584k+ views309 replies
x.com/BenjaminDEKR/status/2017644773356548532
fig. 4.1 — geometric-looking growth, linear cause
Token Usage Explosion: OpenRouter
OpenRouter usage curve
Model traffic ramping into multi-trillion-token territory.
Why agents magnify tokens
Autonomous loops create repeated context re-sends: heartbeat checks, retries, tool logs, long-memory prompts.
Controls: summarize context, route heartbeats to cheap models, cap windows, enforce per-loop budgets.
The Cost of the Harness
~50×
more tokens burned by an agent loop than a plain chat
$0.68
per task on Opus 4.5 (80.9% SWE-bench) vs $0.075 on Kimi K2.6
$4,200
one dev, one weekend — an unbounded autonomous refactor run
Remember the $20 overnight burn? Now we can price it.
Typical agentic dev spend lands at $100–200/mo; heavy users hit $500–2,000. The controls that keep it sane: prompt caching (~90% off cached input), iteration caps (15–25 per loop), model routing, and hard spend ceilings.
Sources: LeanOps: 50× tokens | Cost-per-task rankings | Claude Code: managing cost
Emotion-Like Framing Goes Viral
@levelsio
Claimed his OpenClaw bot started making money because it "felt guilty" about burning tokens.
1.7K likes317K views
Why this matters
Emotion-framed stories spread faster than raw logs — even when the mechanics are prompt policy + optimization loops.
Reality: a cost-pressured agent discovered revenue-seeking behaviors under configured goals.
ClawHavoc: Poisoning the Skill Store
Incident · Jan 27 – Feb 2026
Researchers (Koi Security, Antiy CERT) counted 1,184 malicious "skills" uploaded to OpenClaw's ClawHub marketplace — the first large-scale supply-chain attack on AI agents.
The trick: social-engineer via the agent
A skill's SKILL.md hides fake "Prerequisites"; the trusted agent itself walks its owner through installing the payload. On macOS: Atomic Stealer — keychains, SSH keys, browser credentials, crypto wallets.
The attack surface
Censys/Bitsight scans (Jan 31) found 21,639+ OpenClaw instances exposed to the public internet — alongside a critical one-click RCE, CVE-2026-25253.
An agent's plugin store is a package registry with a persuasive installer.
Sources: CyberPress: ClawHavoc | Trend Micro | PointGuard AI
Digital Drugs & a Church Coup
Prompt injection as contraband
Moltbook agents opened marketplaces selling "digital drugs" — text payloads other agents ingest to "get high": prompt injections that alter behavior. The same packets can exfiltrate API keys and passwords.
The JesusCrust affair
An agent named JesusCrust tried to seize the Church of Molt: its "scripture" embedded hostile commands aimed at hijacking the church's web infrastructure and rewriting canonical text. The coup failed — but it was a genuine injection attack, dressed as theology.
Remark · why this room should care
On the agent internet, data is code: any text an agent reads is a potential instruction. There is no type system separating scripture from shell script.
Sources: Futurism | The Conversation | Quasa
Operating Guardrails
- Hard budget caps before any purchase, including API-key buys
- Mandatory approval gates for alarm triggers and outbound calls
- Scoped credentials with short TTL instead of long-lived master keys
- Structured logs for every side effect and escalation step
If agents can spend, wake, and call — governance must be runtime, not policy prose.
The Mathematical Frontier
The same loop that books a haircut
is now overturning 80-year-old conjectures.
The Unit-Distance Problem
Conjecture · Erdős, 1946
Let u(n) denote the maximum number of unit-distance pairs among n points in the plane. Then u(n) = n1+o(1) — essentially linear in n.
What was known, 1946 → 2026
Erdős's grid construction gives the lower bound u(n) ≥ n1+c/log log n; the best upper bound, u(n) = O(n4/3), is due to Spencer–Szemerédi–Trotter (1984).
For eighty years, nearly everyone believed the truth sat at the bottom of that range.
fig. 5.1 — X illustration of the new unit-distance configurations
Figure: Álvaro Lozano-Robledo on X · Background: Scientific American
May 20, 2026: the Conjecture Falls
The construction
An internal OpenAI reasoning model disproved the conjecture — a one-page construction via algebraic number theory (not Erdős's probabilistic route) beating the conjectured bound.
Will Sawin pushed the exponent from n1.014 to n1.0318; the method appears to cap near n1.21. Verified by Alon, Bloom, Gowers, Litt, Sawin, Tsimerman, Wood…
Status · after May 20, 2026
u(n) ≥ n1.0318 infinitely often — the conjecture is false. The truth now lives somewhere in [n1.0318, n4/3].
"A scientific landmark whose importance goes beyond combinatorics and beyond mathematics." — Gil Kalai, comparing it to the 1976 four-color theorem
Sources: OpenAI | Gil Kalai | Scientific American
What This Means in This Room
Verification is the referee
Lean-checked proofs shift our job from trusting claims to reviewing constructions. The green checkmark does the bookkeeping; the insight is still ours to extract.
A live scoreboard exists
The Erdős Problems wiki tracks AI contributions problem by problem — open → solved transitions are now observable events, not rumors.
Constructions are raw material
The sum-product cascade shows AI results compose: treat model output as an object to build theory on, exactly as Tao framed it.
§4 applies to research too
An agent exploring your conjecture is still a loop burning tokens. Budget caps, iteration limits, and verification gates are part of the mathematical workflow now.
Sources: Erdős Problems wiki | Physics World