
Coding agents now work inside your code
Developers now hand whole tasks to AI coding agents such as Claude Code, Codex, Cursor and GitHub Copilot. The agent reads the repository, runs shell commands, edits files, calls tools and sends code to an AI model, with the developer's own access.
- 84% of developers surveyed are using or planning to use AI tools in their development process, up from 76% a year earlier.¹
- Claude Code-assisted commits showed a 3.2% secret-leak rate, versus a 1.5% baseline across all public GitHub commits. In 2025, 28.65 million new hardcoded secrets were added to public GitHub commits, and 24,008 unique secrets sat in configuration files that connect AI tools to other systems.²
- OWASP's Top 10 for Agentic Applications lists tool misuse, and identity and privilege abuse, where leaked credentials let agents operate far beyond their intended scope.³
- The Model Context Protocol, which coding agents use to reach tools, says tools represent arbitrary code execution and must be treated with appropriate caution.⁴
Every coding agent is a new worker with a developer's access. Argent gives it a name, an owner and rules.
Five questions a security team cannot answer
- Which coding agents are running, and whose are they? A tool on a laptop calls a model with a key, and the key belongs to nobody in particular.
- Where is our AI provider key? It sits in settings files, scripts and pipelines, one copy per agent.
- Which commands did the agent run, and who allowed them? The shell history says what ran, not who decided it could.
- What code and secrets went to the model? A pasted log or
.envfile travels with the prompt. - Which tools can it reach? A new MCP tool appears on a server overnight and the agent can call it.
See AI visibility, data loss prevention, agent identity and ownership and threat defense.
How Argent handles it end to end
Argent Strategic Intelligence sits between each coding agent and what it reaches: the AI model, your MCP tool servers and the actions it takes on the developer's machine. Every agent that works through Argent is checked at each step, and each decision is a signed record.
- A named owner. Each tool is registered as its own agent under an active, named person. When the owner leaves, their agents are refused on the next request. See Identity Tethering and Owner Offboarding.
- No AI key on the laptop. The agent holds only its own Argent key, kept in the OS keychain. Your provider key stays in your vault and the gateway attaches it on the way out. See Keyless Agents.
- Rules for actions. The Argent hook asks the gateway before each shell command, file read or write, web fetch and MCP call. Rules allow it, ask a person or deny it; an action with no rule is denied.
- Secrets out of prompts. Credentials and personal data are redacted before the model sees them, and replies can be checked too. See Sensitive Data Redaction and Response Screening.
- Approved tools and models only. Each MCP tool call needs a scope the agent holds, a new tool on a server the MCP Gate enforces waits for an admin's approval, and only models you allow can answer. See Agent Reach Control, Model Control and Prompt Injection Defense.
- Spend and evidence. Budgets cap what agents spend, sessions can be replayed and the record is signed. See AI Spend Caps, Agent Work Tracking, Session Replay and One-Click Audit Evidence.
A walk-through: one developer, one agent
- Connect. A developer signs in with
argent login, then runsargent connect claude-code. Argent registers a Claude Code agent under her name, stores its Argent key in her keychain and writes Claude Code's settings, with a backup.argent doctorchecks the setup. - Prompt. She asks the agent to fix a failing test. Argent checks the agent and its owner are active and the model is allowed, then attaches the company's provider key from the vault.
- Secret. She pastes an error log that holds a cloud access key. The key is redacted before the prompt leaves for the model.
- Command. The agent asks to run the test suite. A rule allows it. It then asks to fetch a script from a site no rule covers: denied, with the reason
action.deny.no_rule. - Tool. The agent calls an MCP tool it holds no scope for: refused. On a server the MCP Gate enforces, a new tool shows as pending until an admin approves it.
- Record. Every step is a signed record. The session is grouped by Claude Code's own session id, with each sub-agent in its own lane, and an authorized reviewer can replay it prompt by prompt.
- Stop. If the agent misbehaves, an admin quarantines it and its next request is refused; her other agents carry on. If she leaves the company, all her agents are refused on their next request.
The same argent connect command and Argent hook cover Codex, Cursor, GitHub Copilot CLI, Gemini CLI, Kiro, Devin, Cline and OpenCode; each tool's page shows how its traffic reaches Argent.
What the security team sees
| Where | What it shows |
|---|---|
| Agents: Connect tab | Copy-paste setup for Claude Code, Codex CLI, Cursor, the Python SDK and curl, filled with the agent's own gateway, route and key prefix. |
| AI Clients | A catalog of 83 AI clients, from command-line agents and IDE extensions to SDKs, each with its setup guide. |
| Credential Broker | Every vault and secret reference, which agents use it, with their human owners, and its last calls. |
| MCP Gate | Every MCP server and tool your agents can reach, each tool pending, approved, denied or drifted. |
| Skills | Each skill an agent loaded, with its exact version, and whether it is pending, approved or denied. |
| Sessions | One coding session turn by turn, with its sub-agents in separate lanes. |
| Detections | Signals such as a burst of tool calls, a new tool name or a spend spike, with the evidence behind each. |
Agent keys and vault secrets never appear on these pages, only the key prefix and the vault reference. Captured prompts open only for authorized reviewers.
Guardrails built for coding agents
A coding agent runs fast, on a real machine, with real access. These are the rules Argent holds it to.
- Denied by default. An action that matches no rule is denied, and so is an action of a kind Argent does not recognize.
- Fails closed. If the hook cannot reach Argent or gets a bad answer, the action is denied.
- No key in a file. The Argent key lives in the OS keychain and reaches the tool through a key helper or the launcher's environment, never a settings file.
- Pinned for the fleet. IT can deploy Claude Code managed settings that pin the key helper above user and project settings, and push the Argent proxy address and certificate to managed machines with Jamf, Kandji or Microsoft Intune.
- Only reviewed skills. Argent records each skill an agent loads by name and hash, scans it, and with skill enforcement on, only approved versions run.
- Loops caught. Repeated identical calls and retry loops are recorded, and denied in enforce mode.
- Look-alikes flagged. A client that claims to be one coding agent but sends requests shaped like another is flagged.
- Watch first. Each area can run in watch mode, recording what Argent would refuse, before you enforce it. See Watch Before You Enforce and Agent Behavior Watch.
Where it fits and how to start
| For | Why it matters |
|---|---|
| Financial services and insurance | When engineers on payment and policy systems use agents, every command and tool call has a named owner and a signed record. |
| Healthcare and retail | Patient and customer data is redacted from prompts, and agents call only the tools they hold a scope for. |
| Public sector | Run it air-gapped and on-premises, with the record kept on your own systems. |
| Your goals | Data loss prevention, agent identity and ownership, budget and token spend, audit and accountability, compliance evidence. |
| Related use cases | Controlling agent tool access, rolling out copilots and chat safely, AI in regulated and air-gapped environments. |
- Connect. Each developer runs
argent connectfor their tool, or IT pushes the Argent proxy and certificate to every managed machine. Keep the provider key in HashiCorp Vault or OpenBao, AWS Secrets Manager, CyberArk Conjur or Infisical, and people in Okta or Microsoft Entra ID. - Watch. Turn on watch mode and see what Argent would refuse while the work goes on. Shadow AI Discovery finds the agents nobody connected yet.
- Enforce. Switch on action rules, the MCP Gate and secret redaction, area by area.
- Run it where you need it. Air-gapped and on-premises, private cloud or your public cloud VPC. See all capabilities and all integrations.