← AI Catalyst C3All programsHomeSearch
AI Catalyst C3·Core Session - Week 14·1:19:00

Community Session II: The Meta-Harness — Praful Thakkar on AgentGrid, Orchestrator-plus-Workers on One Canvas, Hand-Off Between Models, and OAuth Instead of API Keys

Praful Thakkar Presenter - Outskill member for over a year ('started my journey in 2025 to be more of an AI generalist... today I'm more of a specialist'); solo builder who has shipped ClaudeLMS and a Marketplace listing; runs the AIC Mastermind call · Shivani Host - polls cohorts (C3 with Cohort 2 and 4 guests), relays chat questions, closes · Cody Shah Participant - explains his Hermes + AgentGrid folder-sharing workflow at 0:49

The short version

  1. Praful's arc: 'it moved from tools to agents, to team, to harness, to a meta-harness' - one orchestrator that 'plans, delegates, implements, reviews, tests' across role-specific workers (build, review, Q&A, security), with a human verifier at the end. 'You become the super boss of that power.'
  2. AgentGrid is the demo: every pane can run a different model, a shared 'common context window' pools what they know, the browser and terminal live on the same canvas, and token use shows top-right (he was at 1-3% of a weekly limit mid-demo).
  3. Hand-off is the killer feature for rate-limited subscribers: 'even if you're a $200 plan... three more hours to wait?' - pin the session and continue in another model; the caveat is that Codex 'will not see [the work] in the same eyes' as Claude.
  4. AgentGrid vs Hermes: 'both of them are almost the same' except you can watch AgentGrid work; Hermes 'goes into deep sleep' with only 'dot dot dot' as feedback. The founder plans to put Hermes inside AgentGrid.
  5. No API keys needed if you have subscriptions: /login per vendor, paste the short OAuth code; OpenRouter keys remain an option. Standard tier: Paul is not present; the useful residue is the pattern and the promo code, not the demo itself.

At a glance, three clicks deep

Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.

01The meta-harness: orchestrating harnesses, not running one agentAn orchestrator coordinating role-specific worker agents (build, review, test, security) with a human verif…

An orchestrator coordinating role-specific worker agents (build, review, test, security) with a human verifier; the layer above a single agent harness.

Tools -> agents -> team -> harness -> meta-harness (0:15)

'Set it and forget it, obviously with all the guardrail possible' (0:17)

Roles: build, review, Q&A, security; orchestrator plans, delegates, implements, reviews, tests (0:19)

Human verifier always last (0:26)

Models don't make the harness obsolete (0:30)

↓ Full write-up of this concept

02AgentGrid: many agent panes, a browser and a terminal on one canvas, each pane its own modelA desktop canvas that hosts independently-modelled agent panes plus browser and terminal, so a whole build…

A desktop canvas that hosts independently-modelled agent panes plus browser and terminal, so a whole build stays in one window.

agentgrid.sh, desktop only (0:36)

Settings tour (0:38-0:40)

Project + orchestrator pane, reasoning extra high (0:44)

Per-pane model choice; Sonnet 5 / Codex / Antigravity workers (0:47, 0:53)

Browser pane on localhost:3000; terminal; git optional (0:54-0:56)

Updates every 6-12 hours (0:37)

↓ Full write-up of this concept

03Hand-off: pin the session and continue in another model when a plan rate-limits youSession state is pinned on the canvas so any model can pick it up;

Session state is pinned on the canvas so any model can pick it up; expect interpretation drift between models.

Rate-limit motivation (0:21)

'Pins on the canvas' (0:22)

Interpretation drift caveat (0:22)

Mid-session model switch keeps context (0:47)

↓ Full write-up of this concept

04AgentGrid vs Hermes: watchable versus 'deep sleep'AgentGrid = visible, interactive orchestration;

AgentGrid = visible, interactive orchestration; Hermes = background autonomous agent; share a project folder to use both.

'Almost the same' except visibility (0:28-0:30)

Hermes 'deep sleep', 'dot dot dot' (0:29)

Founder plans Hermes inside AgentGrid (0:27)

Cody's shared-folder bridge (0:49)

Memory tooling: Mem0, MCP, Honcho (0:18)

↓ Full write-up of this concept

05The common context window and what it costsAgentGrid pools context across panes;

AgentGrid pools context across panes; token use is shown live; connectors link workers across projects.

Common context window (0:54)

'Does use a good amount of context' (0:54)

1-3% of weekly limit readout (0:60-0:61)

Filmmaking demo workers; Wixie + ElevenLabs via Claude (0:61-0:66)

Connector between workers question (0:66-0:73)

↓ Full write-up of this concept

06OAuth log-in per vendor instead of API keys; the AICPRO26 promoAttach each vendor account with /login and a short OAuth code;

Attach each vendor account with /login and a short OAuth code; API keys (OpenRouter) optional; six-month Pro via AICPRO26.

/login flow with short code (1:16-1:17)

OpenRouter keys optional (1:16)

Wixie + ElevenLabs; Gemini Flash voice (1:17)

AICPRO26: six months Pro, ~$300 value (0:38-0:39, 1:18)

↓ Full write-up of this concept

The concepts in full

01

The meta-harness: orchestrating harnesses, not running one agent

One agent is a tool. A team with a boss is a harness. A boss of bosses is what he means by meta-harness.

Praful's personal timeline is the framing: prompting, then tools and gems, then agents, then teams, then harness, then meta-harness - 'it moved from tools to agents, to team, to harness, to a meta-harness.' The interface has become 'set it and forget it, obviously with all the guardrail possible', and the point is orchestration: 'we are not going to talk about just one agent, we are going to talk about orchestrated agents, and that is what the meta-harness is all about.' The roles are a film crew's: 'one of them will be building, one of them will be reviewing, one of them will be doing the Q&A, and one of them is going to check the security', coordinated because 'each of them cannot talk to each other, and that's where we need someone who's on top of it' - the orchestrator that 'plans, delegates, implements, reviews, tests', with 'a verifier... the human in the loop will be always there.' Stronger models do not retire the harness: 'stronger models do not make AgentGrid obsolete.'

Why it matters

Same architecture as Paul's own agent-team experiments: the orchestrator is the product, the workers are swappable.

02

AgentGrid: many agent panes, a browser and a terminal on one canvas, each pane its own model

'You don't have to leave this canvas to work on your project.'

AgentGrid (agentgrid.sh, desktop only, Mac/Windows/Linux) is the tool: 'everything can be obtained in this one tool.' Settings expose account and Pro code, themes, API keys, dictation, calling, agents/models, devices, shortcuts. The demo creates a project ('Outskill House demo'), opens a pane with Codex as orchestrator at reasoning 'extra high', then spawns workers - Claude Sonnet 5 for design, Codex for interactive UI, Antigravity for visuals - because 'every pen you open, you are independent of using any kind of a model.' A browser pane loads localhost:3000 for the build preview; 'you can open your terminal window' and 'you can also connect the git' (he skips git because 'I'm right now solo'). The product ships fast and rough: 'they keep on updating every 6 to 12 hours, and I have to keep on installing and restarting.'

Why it matters

A cheap way to try a multi-model team without wiring anything - and a reminder that the tool is still churning.

03

Hand-off: pin the session and continue in another model when a plan rate-limits you

The $200 plan hits its ceiling at the worst moment. Hand the session to the next model and keep going.

'Even if you're a $200 plan, you just feel that, okay, now what do I do? Three more hours to wait?... In AgentGrid... you can just simply hand off the session to any other agent' - 'they call it the pins on the canvas, and then you can simply continue from where you left.' The honest caveat: 'the way Claude looks at the answer or solution... Codex will not see [it] in the same eyes, but again, you can simply switch without having any kind of an issue.' Confirmed in Q&A that a pane started on one model can switch and 'does it share the same context window? Yes.'

Why it matters

Directly relevant to Paul's own weekly-limit juggling between Claude, Codex and the rest.

04

AgentGrid vs Hermes: watchable versus 'deep sleep'

Same job, different visibility: you can see AgentGrid think; Hermes shows you three dots.

'Both of them are almost the same. Except for the fact that... in AgentGrid, you can see you are the boss... we can watch what's happening.' Hermes: 'give the task, Hermes goes into deep sleep, sometimes we wonder whether it's working or not... typing or something, dot dot dot.' From the founder: 'they are also planning to get Hermes within AgentGrid', which 'might be a different product altogether' once out. Cody Shah's workflow bridges them today: 'you give the Hermes your folder of your agent grid, and basically it takes up the same project, and then you can control the same thing from any window.' Memory remains 'one of the biggest hurdles' - Mem0, MCP and Honcho named as the tooling.

Why it matters

Positions the two tools Paul already has notes on (Hermes sprint) as complements rather than rivals.

05

The common context window and what it costs

Five panes do not mean five context bills - but they do mean one big one.

'There's a Common Context Window, which is... common for all of them, so do not get scared by saying that, oh my god, it's going to use so many tokens' - immediately qualified: it 'does use a good amount of context.' Usage is visible live top-right; mid-demo he reads '1%, 3% of the weekly limit.' Workers can also be cross-connected between projects (the 'connector' feature raised by Nadeem), and the filmmaking demo spawned story, screenplay, visual-direction and clip/voice/subtitle workers with Wixie and ElevenLabs reached through Claude.

Why it matters

The number to watch when Paul tries it: one pooled context means one pooled burn.

06

OAuth log-in per vendor instead of API keys; the AICPRO26 promo

'You don't need to have API if you have the subscription to all the models.'

Login flow: 'say slash login. It's going to log into your account, will give you some long URL. Click on that... give you some option to authorize... will give you some 7, 8, 9 character or number key, which you can put it back.' Keys remain optional: 'there was an option on AgentGrid where you can put in your own API keys from OpenRouter, or somewhere else.' Voice: 'voice agent through Wixie. Wixie has 11 labs', and vendors' own voice models ('I used Gemini Flash voice model in one of my voice agents'). The promo: 'I just could manage to get 6 months of free for AgentGrid Pro... unlimited number of canvas and an unlimited number of projects', code 'AICPRO26', 'almost equivalent to $300', also posted on the Mastermind site.

Why it matters

Subscriptions Paul already pays for are enough to try this; no new keys to manage.

Tools referenced

ToolCoverageMomentContext
AgentGriddemonstratedMeta-harness canvas; the whole demo
CodexdemonstratedOrchestrator pane, reasoning extra high
HermesexplainedBackground counterpart; planned integration
Claude CodementionedSonnet 5 design worker; ClaudeLMS built with MCP
AntigravitymentionedVisuals worker; heard 'NP gravity / Agravity'
OpenClawmentionedEarly harness example
GrokmentionedConnectable model
CursormentionedUsed before, not with AgentGrid
DevinmentionedNamed, never used
OpenRoutermentionedOptional key source
ElevenLabsmentionedVoice via Wixie
Mem0mentionedMemory tooling
MCPmentionedMemory / connections; heard 'NAMP'
Wispr FlowmentionedDictation option in settings; heard 'whisper flow'

Session materials

Archived locally on V: — click to open. Companion pages link to the LMS.

Action items

    Resources mentioned

    Resources
    • docAgentGrid (desktop) - agentgrid.sh
    • docSession recording (Google Drive embed)

    Extraction notes

    This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

    Transcript corrections applied

    The transcript saysThe trainer actually means
    off-solidobsolete
    NP gravity / Agravity / Agent VeroxAntigravity
    GPT-6 Astra / Kodak's Estra modelCodex Astra model (the version as spoken is unclear)
    Wynes / VignessVignesh (AI filmmaker trainer)
    NAMPMCP
    codecsCodex

    True on recording day — verify before relying