Cortex crew
Agents that orchestrate AirGPT and Space.
Not a chat window. The crew underneath.
Cortex compiles a goal into a minimal DAG, runs it across AirGPT and Space, and leaves evidence -- SQL, rows, manifest, ledger -- or refuses. Bring your own keys.
WHY CORTEX
Built for execution.
| FEATURE | MANUAL | CHAT / SEARCH | CORTEX |
|---|---|---|---|
| Answers questions | — | ✓ | ✓ |
| Takes actions | Manual | — | ✓ |
| Sandboxed execution | — | — | ✓ |
| Cryptographic audit trail | — | — | ✓ |
| Rollback on failure | — | — | ✓ |
| On-prem / air-gap | Possible | — | ✓ |
| Skill File versioning | Docs | — | ✓ |
THE PROBLEM
Undocumented company knowledge.
SCATTERED
Procedures live in Slack threads, tribal memory, and half-updated wikis. New hires guess. Agents have nothing solid to run.
BLACK-BOX
One giant prompt looks clever until it fails silently. No DAG, no guardrails, no rollback — just vibes in production.
RUNTIME
NL → Skill Files → DAG
Cortex compiles how your company works into versioned Skill Files, then executes them as minimal DAGs you can inspect, approve, and replay.
# refund_handling.skill.yaml triggers: [zendesk.ticket.tagged:refund] knowledge_sources: [kb://refund-policy@v3] tools: [zendesk.respond, stripe.refund] guardrails: max_refund_usd: 500 human_approval_above_usd: 100
PIPELINE
Ingest · Structure · Execute
- Ingest
- 50+ connectors — Slack, mail, CRM, repos, tickets, warehouses — continuous sync.
- Structure
- Skill File compiler extracts procedures, versions them, and waits for human review.
- Execute
- Sandboxed agents, cryptographic audit trail, rollback on failure. Zero-trust by default.
USE CASES
Where Cortex runs
Support refunds
Compile policy into a Skill File. Agent responds, refunds under threshold, escalates above it.
Margin & ops checks
Pull SQL + sheets, flag anomalies, draft the exception report for human sign-off.
CRM triage
Score inbound deals, update stages, hand hot leads to Closer without losing the audit trail.
Compliance audit
Replay every tool call. Export the trail. Prove what ran, when, and under which guardrail.
HOW IT WORKS
Four steps
- 01
Task → DAG
Natural language becomes a minimal execution graph — not one opaque prompt.
- 02
Guardrails & review
Humans approve Skill Files and high-stakes thresholds before agents touch prod.
- 03
Sandboxed run
Tools execute in isolation with scoped credentials and clear failure boundaries.
- 04
Audit + rollback
Every node is logged. Failed runs reverse; successful ones leave a trail you can export.
TRY THIS
Example prompts
- Compile a refund Skill File from our Zendesk macros and Stripe policy PDF.
- Build a DAG that syncs closed-won deals from Salesforce into the weekly ops sheet.
- Flag purchase orders over $10k that skipped two-person approval in the last 30 days.
- Summarize Q3 margin exceptions and draft the board one-pager with source citations.
Run them from AirGPT over the encrypted connect tunnel.
FAQ
Straight answers
How is this different from a public chat tool?+
Chat and search stop at answers. Cortex compiles procedures into Skill Files and runs sandboxed agents that take actions with audit and rollback.
Do you host our data?+
Bring your own keys and infrastructure. Personal operators use AirGPT locally; enterprises can run hybrid or air-gapped. Zero retention is the default posture.
What is a Skill File?+
A versioned, reviewable definition of triggers, knowledge sources, tools, and guardrails — the executable form of how your company actually works.
Can we air-gap?+
Yes. Cortex Runtime is open source. Deploy on-prem, keep connectors inside your network, and govern agents through Suite Admin.
What does a two-week pilot include?+
We connect top data sources, compile your first Skill Files, and demo a live agent run against your actual workflows — no long commitment.
SUITE
Center of the business OS
On Netie Suite, Cortex sits between Vanguard and Closer — streams in, actions out, Admin toggles what may run.
Pilot in two weeks. No commitment.
We connect Cortex to your top sources, compile the first Skill Files, and demo a live agent run against your actual data.