AI transformation · Agents
Silex designs, builds, and governs AI agents and agentic coding environments on Claude, Codex, and open-source agent harnesses, connected to your systems through the Model Context Protocol.
We build with
How agents are governed
Agents reach models and tools only through gateways you control, and a person approves every change until the workflow has earned promotion. Silex applies the same six controls to every agent it builds.
Built on Claude, Codex, or an open-source harness.
Every change waits for a person's approval until the workflow has earned promotion on evidence.
Identity
Every agent and every user signs in through your identity provider before a model or tool call.
Per-tool access
Each agent reaches only the MCP tools that its role allows.
Budgets
Token budgets stop spend at the limit you set for each team or agent.
Redaction
The AI gateway redacts sensitive data before a prompt reaches a model.
Human approval
A person approves every change until the workflow has earned promotion.
Audit
Every call and every approval is logged for review and incident response.
Competencies
Our engineers design the agents, the coding-agent practice around them, and the access, prompts, evaluation, and model routing they depend on.
Silex builds operations agents, for example on the Claude Agent SDK, that act only through approved automation. Each agent pauses for human approval before a change and stays within the budget you set.
Example: our agentic operations demonstration, in which an agent triages an incident and proposes a fix for a person to approve.
Silex rolls out coding agents such as Claude Code and Codex to engineering teams. Access runs through the AI gateway with single sign-on, model access and token budgets by group, shared project instructions, skills, and audit hooks.
Example: the Governed AI Coding Assistants lab, which sets up this rollout for your platform and security teams.
Silex designs MCP servers that give agents well-described access to enterprise systems. We secure them with OAuth, token exchange, per-tool access control, rate limits, and audit logging at the gateway.
Example: an existing REST API exposed as an MCP tool through Kong, with no backend code.
Silex writes system prompts, skills, and output schemas for agents whose results feed automated systems. We test each one against defined criteria before an agent uses it.
Example: agent findings drafted with cited evidence in a fixed format, so a person can review and approve them in one step.
Silex tests models and coding agents on realistic operations scenarios with graded rubrics. Our agent evaluation lab compares Claude, GPT, Gemini, and open-weight models on the same scenarios, with judging across vendors.
Example: a workflow moves up the autonomy ladder only when the agent's diagnoses match your engineers' resolutions.
Silex routes agent model traffic through an AI gateway to the Anthropic API, Amazon Bedrock, OpenAI, and self-hosted models. The gateway applies rate limits, cost controls, and failover to every call.
Example: a self-hosted gpt-oss-120b served behind the Kong gateway alongside hosted models.
Agentic coding practice
When engineers connect coding agents to model providers on their own, each one brings separate credentials, costs, and tool access. Silex rolls out Claude Code, Codex, and open-source coding agents through the AI gateway, so every session uses your identity provider, stays within a budget, and leaves an audit record.
How we engage
Each engagement starts in a lab environment and ends with your team running the agent.
Your engineers build and govern an agent in a lab environment, with the gateway controls in place from the first request.
Silex reviews a planned or existing agent: its tools and MCP servers, approval gates, budgets, evaluation plan, and audit.
The agent runs on your systems behind the AI and MCP gateways, and a person approves every change.
Silex delivers runbooks and moves day-to-day operation of the agent to your team.
Platforms
Certified Delivery Partner
Select Tier Services Partner
Claude Agent SDK · Claude Code · Codex · Open-source agent harnesses · Model Context Protocol · Kong AI Gateway · Amazon Bedrock · Anthropic API · OpenAI API · Self-hosted models on vLLM
Questions buyers ask
We build with the Claude Agent SDK, Claude Code, Codex, and open-source agent harnesses. The choice depends on the workflow, the models you are approved to use, and where the agent needs to run. The gateway controls are the same for all of them.
Agents act only through approved automation, such as Ansible job templates, and never hold credentials for your systems. Every change requires a person's approval until the workflow has earned promotion, and each workflow has a stop control.
Yes. Claude Code, Codex, and similar agents connect through the AI gateway, which signs users in through your identity provider and applies model access and token budgets by group. Every session is logged for audit.
An MCP server exposes tools and data to agents through the Model Context Protocol. Silex puts MCP servers behind an MCP gateway that applies OAuth, token exchange, per-tool access control, rate limits, and audit logging, including for third-party servers you did not build.
We test models and agents on realistic operations scenarios with graded rubrics in our agent evaluation lab, and we compare vendors on the same scenarios. A workflow moves up the autonomy ladder only when the agent's diagnoses match your engineers' resolutions and failure tests pass.
Next step
We review a planned or existing agent: its tools and MCP servers, approval gates, budgets, evaluation plan, and audit.
Claude and Claude Code are trademarks of Anthropic, PBC. Codex and other product names are trademarks of their respective owners.
