how we build

How we build

The tools change every quarter. These are the ones we reach for now, and how we decide between them.

Types of agents we build

Workflow / orchestration — multi-step pipelines with a defined path, branching, and exception handling.
Conversational / assistant — human-facing agents on chat surfaces, tuned for latency and tone.
Background / autonomous — heartbeat, scheduled, or event-driven; they run without a person in the loop for each cycle.
Tool / integration — thin agents that wrap a single system or API over MCP.
Supervisor + sub-agent — a planner that decomposes work and delegates to specialist agents.
Retrieval / knowledge — agents grounded in your documents and data via RAG.

OpenClaw — an open-source (MIT), self-hosted agent framework. A long-lived gateway daemon handles sessions, tool dispatch, and channel routing; a heartbeat drives scheduled and proactive tasks; the ClawHub marketplace supplies ready-made skills; it is model-agnostic. We reach for it when you want a batteries-included agent running around the clock on messaging channels your team already uses. It deploys via its installer daemon or Docker, and its gateway is never exposed directly — only reached over an SSH tunnel or Tailscale.
Claude Agent SDK / custom runtimes — when you need a specific orchestration shape, custom agent graphs, or the agent embedded inside an existing application.
Queue + worker runtimes — for batch and event-driven jobs that don’t need an always-on gateway: a queue, workers, and a scheduler.

Anthropic Claude — Opus, Sonnet, and Haiku — is our default for reasoning and tool use.
GPT, Gemini, Grok, Mistral, DeepSeek — used where a task or a budget favours them.
Local models via Ollama or any OpenAI-compatible endpoint, when data can’t leave your network.
Routing is per task, not per project — the cheap model handles the easy 80%.

We most often build on AWS — Lambda, ECS / Fargate, SQS, EventBridge, Bedrock, Secrets Manager, CloudWatch — it’s what we operate most and it has every primitive an agent needs.
GCP and Azure are first-class: if your team already runs there, we build there, with the same architecture and equivalent services.
Cloudflare Workers for edge logic and x402 metered endpoints.
Everything as infrastructure as code — Terraform, or the platform’s native IaC — with a rollback path.

MCP (Model Context Protocol) is the single boundary between agents and your systems — CRMs, ERPs, databases, internal APIs. It is portable across every runtime above, so swapping runtimes doesn’t mean rewriting integrations.

Auth, sandboxing, spend caps, audit logs, tracing, and evals are on every engagement, not a phase-two add-on. They are enforced in the runtime and tool layer, never as a prompt asking the model to behave.