AI Agents & Bots

Tool use, not prompt engineering

Most "agent" projects we see in practice are bloated prompts shoved into GPT and a hope something comes out. Our work starts with tool design: what the agent can call, with what arguments, what the constraints are, what it returns. It is regular function specification — only the user is a language model.

We work primarily on Claude (Anthropic) for tool-use quality and predictability. Where a cheaper, faster model is needed — GPT-4o-mini or our own open-source deployment on Modal/Replicate.

Memory, state and guardrails

An agent without memory is random. We implement three state layers: short-term (task context, in prompt), working memory (Redis, cleared between sessions) and long-term (Postgres + vector search, persistent).

Guardrails are critical. Every agent has a whitelist of allowed tools, allowed domains, a cost limit per task, a time limit, and human-in-the-loop checkpoints for high-risk actions. Without these, the first bad call turns $200/day into $20,000/day.

Evaluation, not faith

Every agent that goes to production has a representative eval set: 50–500 cases, hand-graded ground truth, success metrics defined with the client. Any change to prompt, model or tool must pass evaluation before it ships. That is engineering practice, not a luxury.

What you get

• Production AI agent with tool use and memory
• Tool definitions and call contracts
• Guardrails system: limits, allowed actions, human-in-the-loop
• Evaluation set and success metrics
• Observability dashboard: cost per task, success rate, latency
• Documentation: prompt, tools, constraints

Stack