Agents built for one job, done reliably.
A harness is everything around the model: tools, memory, routing between providers, guardrails, human handoff, and the evals that tell you it still works after the next model update. I build it from the ground up for your workflow, in your stack.
What you get
- 01 A working agent system
- 02 An eval suite on your real cases
- 03 Documentation and runbook
- 04 A clean handover to your team
Proof
- C++ code ↗
amber
Agent harness in C++ with a terminal UI, inspired by BitchX and ircII. Scored by amber-bench, my 49-scenario harness benchmark; numbers on the work page.
github.com/jtrefon/amber
- Swift code ↗
compass
Native macOS AI IDE in Swift. Agentic coding with your own provider and keys. No Electron.
github.com/jtrefon/compass
- C# private
Off-plan real estate platform, Dubai
.NET core and API, Angular back office, React chat agent, multiple LLM providers via OpenRouter.
- PHP private
Inbound sales agents
PHP agents inside the client’s application, wired to local scripts, SOPs and domain knowledge.
Lab notes on this
- amber-bench Local models for agentic coding, measured.
- 2026-10-27 Agent context management: budgets, truncation, compaction, and knowing when to stop
- 2026-10-27 Agent eval statistics: is 816 vs 808 a real difference?
- 2026-10-27 Constrained decoding: grammars and JSON schema so tool calls always parse
- 2026-10-27 MCP under the hood: JSON-RPC, transports, and capability negotiation
- 2026-10-27 Retrieval math for RAG: embeddings, chunking, hybrid search and reranking
- 2026-10-26 amber-bench: benchmarking the harness, not the model
Scope check
Don’t hire me for this if
You need a chatbot on your website. Buy one.
Have a problem like this?
Describe your problem30-minute technical call with the engineer who builds it. No sales team, no slides.