Viewpoint
What separates an agent that survives production from a demo.
Before an agent can use business data or take action, it needs clear permissions, useful tests, and a way to recover from failure.
These are the checks I use to review an agent design.
02 / What an agent should be
What an agent needs around the model
Planner
Decomposes a request into steps. Reasons over your domain, your tools, and your data.
Tool layer
MCP servers, REST clients, database adapters. Tools have schemas, timeouts, and audit logs.
Memory
Short-term scratch, long-term in Postgres with RLS. No leaking state across users.
Eval harness
Golden tasks, regressions, model upgrade guards. You see quality before customers do.
03 / Stack
Claude. MCP. Postgres. TypeScript.
Anthropic Claude across API, Bedrock, and Vertex. MCP for tool surfaces. Postgres with Row Level Security for state and permissions. TypeScript end to end. n8n when orchestration is the bottleneck. Helicone, OpenTelemetry, or Langfuse for observability.
04 / Multi-agent when it earns it
Hierarchies only after a single agent works.
ClawKeeper explores approval rules for agent actions. CipherClaw investigates failures across multiple agents. Their public repositories let you inspect the implementations and limits.
05 / Proof
Projects you can inspect
My open-source agent projects are on GitHub. I bring 13 years in US enterprise IT, 9 Anthropic Academy course completions, and community support experience as a Lovable Senior Champion.
FAQ
Common questions.
- What is a custom AI agent?
- A program that uses an LLM to plan, call tools, and act on real systems with guardrails, retries, and observability. Not a chatbot.
- Single agent or multi-agent?
- Start single. Add agents only when role separation reduces error rate or cost. Test one agent before introducing coordination between several.
- Which models do you use?
- Claude (Anthropic) as default. OpenAI when tool-use or function quality demands it. Local or open-weight models on request.
- How do you keep agents from going off the rails?
- Evals before launch, guardrails at tool boundaries, Postgres RLS for data access, full logging, and kill switches. The logs should show which action ran and why.
- How do permissions work?
- Agents act as scoped service accounts. Sensitive actions require human approval. Row Level Security in the database enforces who can read or write what.
- Can this integrate with an existing stack?
- Usually. MCP for tool exposure, REST or GraphQL where it already exists, webhooks in both directions. Migration is rarely the prerequisite people expect.
- Where does the paid work live?
- On TechTide AI. This page is the reasoning, not an offer.
Have an agent that almost ships?
Hardening an agent into something your team can own runs through Systems Decision Audit at TechTide AI, where the scope and the pricing live.