pre-product · MVP in 4 weeks · taking 3 design partners→

The brain
your AI agents run on.

Your agents don’t lack intelligence. They lack your company.Nano learns how you actually decide — refund authority, escalation rules, pricing exceptions — and feeds that to every agent on your stack.

Become a design partner ↗Join waitlist →

Every company runs on rules nobody wrote down. Nano serves them to the agent at the moment of decision.

Support agent
refund ≤ $50 — resolve. Above — escalate ops lead.
Sales agent
pricing exception OK if LTV is over $10k. Else — billing.
Incident agent
P0 — page on-call exec within 5 minutes. Comms after.
Recruiting agent
Q4 hiring frozen, except security. Reassess after Jan board.
Not searchNot a wikiNot a chatbotA brain your agents — and your team — run on.
04 — Memory

Memory, not a search bar.

One slice of the brain — a procedure with the variables that decide it. Agents read it as SKILL.md. Humans read it as plain text. The decision view surfaces the variables at the moment of choosing.
your-company/brain/incident-response/SKILL.md
v0.4 · synced 4m ago
Surfaces three ways
The same memory rendered for whoever is looking — agent, human, or the moment of a decision.
For an agent
SKILL.md, MCP-callable. Loaded by the agent on demand.
For a human
Plain text. Open it, edit it, sign off in a PR.
For a decision
Variables surface here when someone has to decide.
Tracked variables · live
incident.severity→sev-2rollback.deploy_id→4f2a91ic.last_assigned→@mayaexec_paged.threshold→$$$ revenuerca.template→v3
05 — The loop

Watch. Extract. Remember. Surface.

Every source changes. So does the answer. The loop runs again — re-extracts, re-remembers, updates the variables that decide things.
00

Watch

Connect the places knowledge lives. Meetings, Slack, tickets, runbooks, on-call notes, customer calls, email.

watching:
  · #ops, #payments, #cx-help
  · zendesk, linear, pagerduty
  · gmeet transcripts, runbooks/
00

Extract

Pull out the variables — the things that actually decide what happens. A timestamp, a person, a threshold, a fix that worked.

variables tracked: 384
  decisions logged: 71 this week
  unresolved contradictions: 3
00

Remember

Put them on a knowledge graph with temporal state. Procedures, decisions, fixes, connections. Every change keeps the lineage.

graph:
  ▢ procedures · 38
  ◇ variables · 384
  → decisions · 1.2k
00

Surface

When someone — or some agent — has to decide, the brain returns the relevant variables, the procedure, and what was decided last time.

surfaced for: @maya · 02:14
  procedure: incident-response
  vars: severity, last_RCA
06 — In the loop

Where the agent actually decides.

Nano sits in the decision layer. Watches where calls get made, learns the variables that move them, then answers the agent at the moment of choosing — with the rule, the threshold, and the precedent.
01Incident response

How engineers respond to incidents.

SOURCE · raw
alerts-bot02:14🔴 PagerDuty: api-gateway p99 > 4s · sev2
@maya02:15i’ll IC. opening #inc-7821
@dev-lin02:16rolling back deploy 4f2a, comms in 2
@maya02:18STATUS: mitigating · ETA 5m · cust impact: low
@dev-lin02:23rollback complete. p99 back to 280ms. opening RCA doc.
NANO REMEMBERS · incident-response
name: incident-response
description: Triage, mitigate, RCA.

## Procedure
1. Open `#inc-<short-id>`.
2. Assign IC + comms + scribe.
3. Mitigate first. RCA after.
4. IC posts STATUS every 10 min.
02Agent at a decision

An agent has to decide. Nano answers.

SOURCE · raw
support-agent · v314:21ctx: customer acme-co, plan=growth, LTV $14.2k, churn-risk=high
support-agent14:21intent: refund $180 on invoice INV-9042 — failed charge ×2
support-agent14:21→ skill_call("refund-authority", { amount: 180, ltv: 14200 })
nano · brain14:21← decision.allowed=true · authority=ops_cap · approver=@ops-lead
support-agent14:22executed. logged to graph. customer notified.
NANO REMEMBERS · refund-authority
name: refund-authority
called_by: support-agent · v3
decided_at: 14:21:04

## Variables (resolved at call time)
- support_cap     = $50
- ops_cap         = $500 (since LTV > $10k)
- escalate_above  = $50

## Decision
amount=$180 → ops_cap branch → allowed
approver=@ops-lead · async-notified
post-action: write decision back to graph
07 — Auditable

Every change is a commit.

The brain is text. Diff it. Roll back. Approve it. Audit who changed what when, and which variable shifted because of it. The thing your agent reads is the same thing your team can sign off on.
your-company/brain/incident-response/SKILL.md
4f2a91clay@4m ago
1@@ skills/incident-response/SKILL.md @@
2 name: incident-response
3-description: Triage and resolve production incidents.
4+description: Triage, mitigate, RCA. IC posts STATUS every 10m.
5
6 ## Procedure
7 1. Open `#inc-<short-id>`.
8 2. Assign IC + comms + scribe.
9-3. Page on-call EM.
10-4. Mitigate first. RCA after.
11+3. Mitigate first. RCA after.
12+4. If revenue impact suspected, page on-call exec.
08 — Any caller

One brain. Many callers.

The brain doesn’t care who’s asking. An agent on a runbook task. A founder in a planning doc. A new hire trying to find the right macro. Same memory, same variables, rendered for whoever called.
CENTER
your-company / brain
Owned. Versioned. In your control.
EDGES
same wire, both ways
Open protocols agents already speak — and a query interface humans can use directly.
NODES
agents, founders, employees
You don’t pick the caller. You pick the procedure. The brain answers either way.
09 — The standard stack

We don’t replace your stack. We slot into it.

The agent stack already has standards. Nano is the missing layer that turns your company’s knowledge into one of them.
Inter-agent comms
A2A
How agents discover each other and delegate work. Donated to LF AI & Data.
Company memory
Nanoyou are here
How a company remembers procedures, decisions, fixes — and the variables that decide what. — Nano lives here.
Procedural rendering
Agent Skills
How a memory is rendered for an agent task. SKILL.md spec, open since Dec 2025.
Tool / data access
MCP
How an agent calls APIs, databases, file stores. JSON-RPC over HTTPS, donated to the Linux Foundation.
Commerce
AP2 / UCP
How agents transact on behalf of a user.
Governance
AEGIS · AgentOps
Guardrails, audit, eval. Where the operations team lives.
What we’re building it on
graph + vectors SurrealDB · temporal state, hybrid retrieval
extraction fine-tuned Gemma / Qwen for variables, contradictions, diffs
synthesis Claude Opus + GPT for reasoning that needs depth
surface SKILL.md over MCP · Python + TypeScript
10 — Why it matters

What was tribal becomes the brain.
The one your agents run on.

Every company has a way of doing things. It lives in three engineers’ heads, a 2019 Slack thread, a macro nobody touches. The agents you ship don’t see any of it — that’s why they fail. The model is fine. It just doesn’t know that this customer is on legacy pricing, that the refund policy changed last sprint, that this incident type escalates differently. Nano writes it down. Every procedure, every variable, every decision. The brain remembers, so your agents — and your team — don’t have to guess.

— the team building Nano
Built by

Lay Sheth. Solo technical founder. AI engineer at Neosapien, software engineering specialist at x.ai. Cut production LLM inference latency 12× at Neosapien. Merged PRs in Zed and rustup. SurrealDB technical ambassador. Writing every line of Nano.

GitHub ↗LinkedIn ↗Site ↗

Three design partners. Eight weeks.

If your team ships agents that don’t know how the company actually decides things — refund authority, escalation rules, pricing exceptions — this is for you. We’re taking three design partners before closed beta.

Become a design partner ↗Join waitlist →Read the manifesto →
© 2026 Nano · made with too many late nights
pre-product · the missing layer