Your developers use agentic tools — or should be.
TL;DR. Claude Code, Cursor, GitHub Copilot Workspaces, Codex — autonomous agents write real code. Productivity rises, security risk too. LLM gateway with policies, ADR onboarding, read-only prod, audit trail, RL scoring.
What this is about
Developers already use Claude Code, Cursor, Aider, GitHub Copilot. Some in private repos, some send code to OpenAI or Claude API. Controlled looks different. AI agents deliver measurably higher code quality, but security control disappears.
An LLM gateway (dyn.ai-core) runs in your EU cloud or on-prem. All coding-agent requests flow through it — API keys centrally managed, no credentials in repos. Policies define which LLMs (Claude, GPT-4, Mistral), budget caps per squad, code-review integration, read-only access to production code, write access only to feature branches. Agents see what you allow. You see what agents do.
How we run it
Nova9 modules: LLM Gateway (central control, policy engine), Business Agents (repository onboarding with ADRs and INDEX files), Observability (telemetry — which squads use which tools, which LLM, how much code output). Gateway runs centrally, agent clients (Claude Code, Cursor) connect with bearer token. Each agent gets sandboxed access — read-only production, write to feature branches, peer review enforced. You can revoke immediately if a rogue agent is found.
When it fits
Tech companies with 10–200 developers wanting AI-assisted coding but not uncontrolled. Startups scaling fast. Finance sector where code audit is standard. Regulated industries where "we don't know which LLM generated code" isn't acceptable.
- SaaS startups (rapid iteration, security non-negotiable)
- Financial services (audit trail enforced)
- Public administration (traceability)
- Enterprise with decentralised IT (governance without blockade)
What we don't do
We don't write code ourselves — those are agent tools. We don't build domain-specific copilots. We don't buy your code history. All data flows through your gateway — you own it. BYO-LLM-key is standard — you control which models are paid.
What you can hand off
-
Tool selection and sizing
Per squad: Does your frontend team need Claude Code or is Cursor cheaper? Business case + licence management.
-
Gateway integration with policies
Budget caps, DLP rules (no production secret over gateway), review enforcement, model whitelist.
-
Repository onboarding
ADRs (Architectural Decision Records), AGENTS markers, INDEX files in repos. Agents can use context.
-
Sandboxing and permissions
Read-only production, write only feature branches. Peer review before merge enforced.
-
1-day training per squad
How to use agents safely? What's the workflow (Agent → Draft → Review → Merge)?
-
Telemetry + RL scoring + policy updates
Dashboard: which squads, tools, LLMs. Reinforcement learning on false-positive reviewers. Monthly policy tweaks.
Engagement facts
| Typical engagement size | 10–200 developers · 3–15 squads |
|---|---|
| Setup phase | 2–3 weeks gateway deployment + repo onboarding |
| Deployment model | EU cloud or on-prem · BYO-key standard |
| Minimum term | 12 months · 3-month cancellation to quarter-end |
| Supported agents | Claude Code, Cursor, Aider, GitHub Copilot Workspaces, open-source agents |
| LLM agnostic | Gateway abstracts — Claude, GPT-4, Mistral, locally-run models |
| Code ownership | 100% yours — gateway logs, doesn't persist |
| Clone handover | Optional · from day 30 · gateway config and policies as code |
Asked before the briefing
-
Which agents do you support?
Claude Code, Cursor, Aider, Codex, GitHub Copilot Workspaces, open-source agents. Unlisted agents possible via API integration. -
Do our repos and code leak?
No. Gateway runs on-prem or in our EU cloud. BYO-key means: you send the API key to OpenAI/Anthropic, no router in between. No training data sharing. -
What if the tool landscape changes?
Gateway abstracts. If Cursor gets replaced, change config, no re-onboarding. API stability is the lever. -
Who pays for LLM usage?
You — via your subscription with the LLM provider. We manage keys and policies only. -
Can agents access our databases?
No — that's an architecture decision. Agents see repos and code. DB access is other tools (business agents, not coding agents).
Gateway and first squad in 3 weeks.
We deploy the gateway, onboard a pilot squad (frontend or backend), write policies, hold training. Then you evaluate: productivity high enough? Then roll out to more squads.