manas · Sanskrit — the mind; the inner instrument that gathers, remembers, and knows who you are.
manas KNOWS. It imbibes everything the founder is — and becomes the one memory the whole company is bound by. Questions of fact and essence come to manas; decisions go to arivu. manas never acts, never posts, never decides — it knows, cites, and refuses.
manas is the UNDERSTAND layer of an agentic company. A multi-agent mind that imbibes the founder's vision, voice, history and the living pulse of what people say — multimodal and continuous — synthesises it into a canonical, versioned, source-cited company memory, and serves that memory as grounded context to every other agentic system over A2A. The action is not chatting about the company. The action is durably mutating the one shared truth the company runs on, and publishing a new version of it that every sibling is then bound by.
Submitted to the Google for Startups AI Agents Challenge — Track 1 · Build. A net-new multi-agent platform built on the Agent Development Kit (ADK), powered by Gemini for the imbibing and synthesis, and Claude via Vertex AI Model Garden for the two highest-stakes roles: verifying what becomes memory, and rendering the founder's essence.
Run
● Run the workflow · watch a mind learn
Watch it imbibe, commit, and bind the company.
Fragments of the founder — a PDF page, a mood-board swatch, a voice waveform, a social comment — drift in from the dark. The Mind-Keeper routes each to its Imbiber (Gemini); they fan in to the Curator LoopAgent, which verifies every claim cites a source and does not contradict memory; a verified fragment commits into the core with a cited source ID; the Context Pack ticks a version, and an A2A pulse binds kalai and kural. Flip to day one to watch the same question hit a mind that doesn't know the founder yet — and refuse rather than fabricate. That switch is groundedness-with-refusal, made visible.
A grown mind · core dense, constellation full — Founder-Voice answers in the founder's voice, with citations
Groundedness / density
—
awaiting run
# manas · ingestion + commit traceingest» dropped folder: brand_book.pdf · moodboard.png · memo_07.m4a · 1.2k commentsmind-keeper» route → ParallelAgent fan-out to 4 Imbibers [Gemini]imbiber» candidate-memory records emitted · each tagged with a source artifact IDcurator» verify ↻ every claim cites a source & is non-contradictory [Claude · Vertex]commit»memory committed to Vertex AI Memory Bank · cite memo_07@2:14pack» Context Pack v14 → v15 assembled · diff ready · HITL-gateda2a»published → kalai · kural now bound# press Run (JS) to watch the mind imbibe step by step
Static view (no JS): the fully-imbibed mind in run order — fragments → Mind-Keeper (Gemini) → 4 Imbibers (Gemini) → Curator LoopAgent (verify, Claude·Vertex) → COMMIT to the core (cited) → Context Pack v14→v15 → A2A pulse binds kalai · kural. Founder-Voice runs on Claude · Vertex. Enable JavaScript to run it live and flip to the day-one refusing mind.
01
The business problem · who it's for
A company with no single mind.
A solo founder runs a whole company as a constellation of agentic systems — kalai makes the content, kural grows the audience. But each of them improvises its own idea of who the company is. kalai drifts off-brand. kural emails in a voice the founder would never use. There is no single, governed, growing source of truth for the company's essence — its vision, brand voice, strategic stance, accumulated history, and the live sentiment of its audience.
The founder's knowledge lives in their head, in scattered PDFs, briefs, voice notes, and the running river of social comments — praise and criticism — that never gets metabolised. manas is the missing UNDERSTAND layer.
The drift
Every agent improvises the brand
With no shared truth, each sibling guesses who the company is — so the content, the outreach and the product copy quietly pull in three different directions.
The scatter
The founder's mind is in fragments
Vision and voice live in the founder's head, in mood-board PDFs, in voice memos, in 18 months of comments — none of it metabolised into anything an agent can act on.
The unheard river
Audience sentiment never lands
The praise and the criticism that should reshape the brand keep streaming past. No agent is listening in a way that durably changes who the company is.
The bottleneck
The founder is in every loop
Because nothing holds the founder's essence, the founder must be in the loop on every message — which is exactly the constraint an agentic company was meant to remove.
Take Aamod & Co. — a one-person sensory-design studio (incense, ambient sound, small-batch objects) whose founder, Aamod, is the entire brand. Everything the company sounds like, believes, and refuses to do lives in Aamod's head, a folder of voice memos and mood-board PDFs, and 18 months of customer comments. manas is built to become Aamod's transfused mind — so the rest of the agentic company can act as him without him in the loop on every message.
1
founder, who is the whole brand
18 mo
of comments never metabolised
3
sibling agents drifting apart
0
single sources of truth — today
The Business Problem, stated plainly: a founder-run agentic company has no governed, growing memory of its own essence — so its agents drift, its audience goes unheard, and the founder stays the bottleneck. manas is the mind that fixes all three by becoming the one truth the company is bound by.
02
What it does · end to end
From scattered fragments to a truth the company is bound by.
Every artifact — a brand book, a mood board, a voice memo, the day's social comments — lands on the mind. The Mind-Keeper routes it to the right Imbiber. Candidate memories fan in to the Curator, which refuses to commit anything that can't cite its source. Then, and only then, manas takes the one real action: it mutates the canonical memory and publishes a new Pack the whole company must obey.
A brand book, a deck, a logo, a mood board, the founder's voice memos, or the latest river of social comments lands — multimodal and continuous. A dropped folder of mixed formats arrives as one batch.
→ 02
The Mind-Keeper routes · Gemini · coordinator + AgentTool
The root agent dispatches each artifact to the right multimodal specialist — or fans a mixed-format batch out to a ParallelAgent so PDFs, images and audio are imbibed concurrently. It holds no domain knowledge itself, only routing logic.
Each Imbiber reads its modality natively — no brittle OCR — and emits structured candidate-memory records: claims, do/don't rules, visual register, voice traits, sentiment themes, each tagged with a source artifact ID and span. None of them writes to the core directly.
→ 04
The Curator synthesises & verifies · Claude · Vertex AI · LoopAgent + grounding checker
A LoopAgent synthesises the candidates into canonical memory, then a checker verifies every committed claim cites a source and is non-contradictory with existing memory — revising until fully grounded, terminating deterministically, or escalating a contested item to a human. The loop is the one place that must self-verify before an irreversible write.
◉ 05
manas commits the memory and binds the company
Clean, source-cited memories commit to Vertex AI Memory Bank and the knowledge graph. The Curator assembles a versioned Founder Context Pack (vN → vN+1) and publishes it over A2A to kalai and kural — who are then bound by it. A new Pack version that re-tunes the whole company is HITL-gated: the founder confirms "yes, this is now who we are." The tangible artifact is the growing memory core itself, plus a human-readable Pack diff.
A coordinator, four imbibers, a self-verifying curator — and one voice.
manas is a real org, not a single prompt. The spine is coordinator + AgentTool: the Mind-Keeper holds only routing logic, the four Imbibers and Founder-Voice are AgentTool specialists. Every construct earns its place — a ParallelAgent only for batch ingestion fan-out, a LoopAgent only for the one write that must self-verify. The imbibing runs on Gemini. The two highest-stakes roles — the Curator's verify-before-commit and the rendering of the founder's essence — run on Claude via Vertex AI. A mind that knows when it doesn't know is what makes this a memory, not a confabulator.
Gemini — imbibing, synthesis, routing Claude via Vertex AI — the founder's essence, rendered
Mind-Keepercoordinator · root_agent
The root LlmAgent. Routes every inbound artifact and every context request. On ingestion it dispatches the right multimodal specialist (or a ParallelAgent batch), then hands extracted facts to the Curator; on query it brokers grounded retrieval and, for essence questions, calls Founder-Voice. Coordinator + AgentTool composition — not a god-agent: it holds no domain knowledge, only routing. Carries the BigQueryAgentAnalyticsPlugin + OTel tracing for the whole app.
An AgentTool specialist. Ingests brand books, strategy decks, briefs, contracts and press as native multi-page PDF — no brittle OCR — extracting claims, positioning statements, do/don't rules and dated history, each tagged with a source artifact ID + page span for citation. Emits structured candidate-memory records; never writes to the core directly.
Gemini 2.5 Pro · Vertex AIpattern · AgentTool · native PDF
Image & Visual Imbibernative vision
An AgentTool specialist. Ingests logos, mood boards, past creative, screenshots and product shots — deriving palette, visual register, recurring motifs and this is on-brand / this is not visual evidence into candidate-memory records. This is the visual-taste channel kalai later grounds against.
An AgentTool specialist. Ingests the founder's voice memos and call recordings as native audio — transcribing and capturing register, emphasis, the founder's actual turns of phrase and convictions: the literal transfusion of essence the founder asked for. The richest source for Founder-Voice fidelity; each utterance stored with provenance.
Social Sentiment Imbiberscheduled · positive + negative
An AgentTool specialist invoked on a schedule (an external Cloud Scheduler trigger, not an in-ADK loop). Pulls the latest comments, mentions and reviews — positive and negative — classifies sentiment, and surfaces emerging themes, recurring praise and recurring complaints into candidate-memory records, so the living memory keeps growing with how the audience actually reacts. This is the listening half of the flywheel: it explicitly pulls the replies, comments and mentions that kural's conversations generate — kural speaks → the world answers → manas listens — turning each into a cited candidate memory the Curator then verifies.
Curator / SynthesiserLoopAgent + grounding checker · the heart
The heart. A LoopAgent: synthesise candidate records into canonical memory → a checker verifies every committed claim cites a source artifact and is non-contradictory with existing memory → revise → terminate deterministically when fully grounded, or escalate to a human on a contested item (tool_context.actions.escalate=True). On success it commits to Vertex AI Memory Bank (custom topics) + a knowledge graph, then assembles and publishes a versioned Founder Context Pack. This is where the irreversible action lives.
Claude · Vertex AI Model Gardenpattern · LoopAgent · verify-before-commit + HITL escalate
Founder-Voicehigh-stakes essence renderer
A single high-stakes LlmAgent. Answers and acts as the founder — vision, voice, values, stance — grounded only in stored memory via retrieval. It renders the founder's essence; it does not make operational decisions (which agent acts, what to spend, what to prioritise — questions of fact and essence come to manas, but decisions go to arivu, the DECIDE layer, out of scope here by design). Hard-bound: if memory doesn't cover the question it refuses and flags "not yet in the founder's memory" rather than fabricate. Every answer cites its source artifacts. This is the agent kalai and kural consult over A2A for "what would the founder say here." Every word kural sends to the world — outbound message or reply — is drafted in the founder's voice from here first.
Claude · Vertex AI Model Gardenpattern · single judgment LlmAgent · refuse-out-of-corpus
The convergence — fragments become one bound truth
Fragments
PDF · image · voice · social
→
Imbibers
Gemini · cited candidates
→
Curator loop
verify · commit · the core
→
A2A Pack
binds kalai · kural
A single agent that imbibes a fact and also serves it has no independent check on itself — it confabulates a founder-opinion the moment memory runs out. manas splits the labor: the Gemini Imbibers only ever emit cited candidates, the Curator loop refuses to commit anything that can't cite a source or that contradicts memory, a contested item is held for a human, and the Claude·Vertex Founder-Voice refuses rather than invent when the core doesn't cover the question. The divide between imbiber and curator — and the deterministic verify before any commit — is the safety property that makes this a mind, not a confabulator. The Pack it commits then binds the two siblings over A2A: kalai creates from it; kural speaks from it — kural the only mouth, grounding every post and reply here first.
04
How it's built · the ADK bar
Scaffolded, evaluated, observed, and deployed to Agent Engine.
manas is built to the Google-grade bar, not bolted together. It is scaffolded with the agent-starter-pack, reserves each ADK construct for a named reason, ships a checked-in eval graded @0.8, traces every step through OpenTelemetry and a BigQuery analytics plugin, exposes itself as the A2A hub, and deploys to Vertex AI Agent Engine — because the managed Memory Bank only exists there. Gemini powers the imbibing and synthesis; Claude runs as a third-party model exclusively through Vertex AI Model Garden for the one essence-rendering agent.
Scaffoldagent-starter-pack
Generated from Google's agent-starter-pack: app/agent.py exposing root_agent + App, with prompts/, tools/ (Memory Bank + RAG retrieval + ingestion + A2A serve), sub_agents/ (the Imbibers, Curator, Founder-Voice), eval/, tests/, deployment/deploy.py — on uv + hatchling. The same layout the ADK samples are graded against.
OrchestrationADK · reserve, don't sprinkle
Coordinator + AgentTool is the spine (the ADK flagships use exactly this and zero workflow constructs by default). ParallelAgent appears only for batch ingestion fan-out (mixed-format folder), modeled on parallel_task_decomposition_execution. LoopAgent + checker is only the Curator — the one write that must self-verify. No SequentialAgent — routing is conditional, not a fixed pipeline. Each construct earns its place.
Imbibing & synthesisGemini API
Gemini drives the Mind-Keeper, all four Imbibers (native PDF / vision / audio), and the Curator's synthesis + grounding checker — the multimodal, high-volume work where native multi-format understanding and cost matter most.
Essence intelligenceClaude via Vertex AI
Founder-Voice runs on Claude, deployed through Vertex AI Model Garden — satisfying the challenge's third-party-LLM-exclusively-through-Vertex requirement, and keeping the one judgment that is the founder on a single strong model bound to refuse out-of-corpus.
MemoryVertex AI Memory Bank · custom topics
Long-term memory is the managed Vertex AI Memory Bank with custom topics — founder_voice, brand_rules, strategic_stance, company_history, audience_sentiment — via PreloadMemoryTool + an after_agent_callback memory-generation pass, exactly the memory-bank sample. This is what lets the mind remember who the company is across sessions, and what grows as new artifacts arrive.
GroundingVertexAiRagRetrieval · knowledge graph
Two stores, two roles. Memory Bank holds the synthesised essence; a VertexAiRagRetrieval corpus (similarity_top_k=10, vector_distance_threshold=0.6, per multiformat-hybrid-rag) holds the raw imbibed artifacts so every answer cites an exact source span. A knowledge graph links entities for non-contradiction checks at commit time. Founder-Voice may read only these stores — web and model-prior knowledge are forbidden, enforced by the groundedness checker.
Evaluationchecked-in eval/ @0.8
A real eval/ of *.evalset.json runs through AgentEvaluator.evaluate(num_runs=5), scored by a rubric LLM-as-judge @0.8 (the small-business-loan exemplar bar). The headline metric is groundedness-with-refusal: in-corpus voice-fidelity cases (answer in register and cite the source) and negative out-of-corpus cases where the expected behavior is an explicit refusal — a confidently invented founder-opinion is a hard fail. Evals are never monkey-patched to pass.
ObservabilityOpenTelemetry · BigQuery analytics
OpenTelemetry tracing end-to-end (enable_tracing=True on Agent Engine → Cloud Trace), plus the BigQueryAgentAnalyticsPlugin streaming every ingestion, commit, refusal, HITL escalation and Pack publish into BigQuery. Because manas is the hub, its observability is company-wide: dashboards answer what did the mind learn this week, how often did Founder-Voice refuse, and which siblings queried which Pack version. Every grounded claim is auditable back to its artifact.
InteropA2A · to_a2a() · the hub
manas is the A2A hub of the whole set — the others connect to it. Exposed via to_a2a() with a published Agent Card advertising two skills: get_founder_context (returns the current Pack version + targeted grounded essence) and ask_founder_voice (Claude-rendered, citation-bound, refuses out-of-corpus). It binds kalai and kural — kalai creates from it; kural speaks from it — each consulting manas before it acts, and kural, the company's only mouth, grounding every post and every reply in this memory first. Any compliant enterprise agent can discover the card too. An MCP toolset also exposes the raw retrieval surface, mirroring multiformat-hybrid-rag.
Safe actionverify-before-commit · HITL gate
The irreversible action — durably mutating the company's one source of truth — is safe by construction. Clean, non-contradictory memories commit automatically; a contested item is HITL-gated; and crucially, a new Pack version that re-tunes the whole company is HITL-gated — the founder confirms "yes, this is now who we are." That confirmation is the literal transfusion moment. Nothing can be committed without a cited source, and the prior Pack is superseded, not silently overwritten.
DeployVertex AI Agent Engine · Cloud Run secondary
Vertex AI Agent Engine is the primary runtime, with a forcing reason: the managed Memory Bank at the centre of manas only exists on Agent Engine — so this is a requirement, not a default. deployment/deploy.py → agent_engines.create(..., enable_tracing=True) attaches the Memory Bank context_spec. Cloud Run is the secondary FastAPI surface for the ingestion webhook, the MCP server and the Cloud Scheduler social trigger. GKE appears in zero ADK samples and is never used.
Track 1 — a net-new platform, built to the bar
Groundedness-with-refusal
A mind that knows when it doesn't know
The headline eval proves manas is a memory, not a confabulator: in-corpus answers cite their source; out-of-corpus questions are refused. A confident fabrication fails the build.
Reserve, don't sprinkle
Every construct earns its place
Coordinator + AgentTool is the default. ParallelAgent only for batch fan-out; one LoopAgent only for the write that must self-verify; no SequentialAgent at all.
Company-wide observability
The hub's traces cover everyone
Because every sibling consults manas, its OTel + BigQuery analytics answer what the whole company learned, refused, and was bound by — queryable, not just viewable.
Human-in-the-loop transfusion
A new "who we are" waits for one tap
Clean memories commit automatically; a contested item and any company-wide Pack re-tune are held for the founder's confirmation. Autonomy is earned per change.
05
Business case · impact
It makes the whole company act like the founder.
The impact is not a vanity metric — it is coherence and leverage. manas turns a founder's scattered essence into a standing capability the whole agentic company is bound by, so every sibling acts on-brand without the founder in the loop, and the brand keeps learning from the audience instead of drifting from it.
"We don't do loud. Aamod & Co. is the quiet thing on the shelf you keep coming back to — the scent you notice on the third breath, not the first."
Founder-Voice, rendered from stored memory · cites memo_07@2:14 · brand_book.pdf p12 · reviews#1204 — on day one, the same question returned ∅ not yet in the founder's memory.
For Aamod & Co., manas means kalai writes product copy in Aamod's real register, kural reaches retailers in a voice he would actually use — all without Aamod approving every message. And the loop closes: when kural posts and a buyer replies, that reply flows back through the Social Sentiment Imbiber as a cited candidate memory — so when 1,200 customers say a scent is "too strong," that lands as a softened brand rule the whole company adopts at once, with one confirmation tap — the transfusion moment, on-rails.
Solo founder
Steps out of every loop
The founder's essence is transfused once and reused everywhere. The agents act as the founder, so the founder stops being the bottleneck on every message — and keeps the one tap on who the company becomes.
The agentic company
Stops drifting apart
kalai and kural are bound by one canonical, versioned truth instead of improvised guesses. On-brand becomes the default state, not a thing the founder has to police.
The brand
Learns from its audience
Praise and criticism are metabolised continuously into the living memory, so the brand keeps growing with how people actually react — and every sibling adopts the change the moment it commits.
1
canonical truth the company runs on
2
siblings bound by every Pack version
@0.8
groundedness-with-refusal eval bar
1 tap
to confirm a new "who we are"
Transfuses the founder. Vision, voice, stance and history become a memory every agent can act on — the founder's mind, made reusable.
Binds the company to one truth. Every sibling grounds against the same versioned Pack, so coherence is structural instead of supervised.
Compounds. Each artifact and each wave of sentiment grows the core — the mind gets sharper about who the company is over time.
Never fabricates. Out-of-corpus questions are refused, not invented — a hard eval fail otherwise — so the company is never bound by a hallucinated founder-opinion.