AI Agents & Harness Engineering
Agents that finish the work, on harnesses you can measure.
We build custom AI agents on proven harnesses (Claude Code, Pi and the major agent SDKs), on cloud models or private local LLMs. We don't reinvent the harness; we configure what is specific to you and measure whether each change helped.
What we configure and measure
The same model can be unreliable or excellent depending on the system around it. That system is the harness. Teams that work on harnesses full time build excellent ones, so we start from a proven harness and spend our effort on the parts that are specific to your business.
Instructions and context
What the agent sees at each step: rules, examples, documents and what it must not do.
Retrieval and memory
Finding the right documents and keeping what matters across a long task.
Tools and permissions
Scoped access to your systems, with the least power the job needs.
Routing and models
Which model handles which step, including cheaper or local models where they hold up.
Verification and recovery
Checks before an agent acts, approvals that bind to the action, and safe handling when a step fails.
Execution controls
Budgets, spending limits, cancellation and stop controls that actually stop the agent.
Ways to work with us
1. Start from a proven harness
Claude Code, Pi or an agent SDK, chosen on your tasks rather than on a vendor benchmark.
2. Configure what is yours
Instructions, skills, tool permissions, routing to cheaper or local models, and verification steps that encode how your team works.
3. Measure every change
Before/after runs on a fixed test suite, one change at a time, so you keep what helps and drop what only adds cost.
Custom agent builds
One workflow at a time: document triage, reporting, data exploration, internal support. Cloud models or local LLMs when data cannot leave your network. Typically 3 to 6 weeks for a first production agent.
Agentic AI & context engineering →AI Analytics Agent Build
Ask your data questions in plain English and get answers, charts and the SQL it ran, with an evaluation set of real questions. 4 to 8 weeks.
See the package →Harness Configuration & Cost-Down
Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.We configure and extend a proven agent harness for your work (instructions, skills, tools, permissions, routing, verification) and measure the result. We change one thing at a time and compare before and after, including whether a cheaper model holds quality.
Scope and limits →Model & Harness Comparison
Pilot: Procedure designed and deliverable with existing tools, but not yet delivered to a client. Sold under pilot terms.Which harness and model should you standardize on? We run the candidates and your current setup on your own tasks, with repeated trials, and report quality, cost and failure modes. We have no stake in which one wins.
Scope and limits →Agentic engineering enablement
Teach your agents how your team engineers: skills, rules, review gates and evals inside the harness your developers already use, with hands-on training so the practice stays after we leave.
Training details →TrendLab Improve
Roadmap: Not deliverable yet. Never quoted or included in a statement of work.An improvement loop that proposes agent changes, with an evaluation process the loop cannot modify. It runs internally on TrendLab's own agents today and is not yet offered to clients.
Our research harness
TrendLab built and runs its own experimental agent harness as a research lab, to learn from the inside where planners, verifiers and model routing help and where they only add cost. It informs how we configure and measure harnesses for clients. We do not sell it.
Building and evaluating are kept apart
When we build or change an agent, our tests of it are internal evaluations, never called independent and never given an Assurance Verdict. If you want third-party evidence about an agent we built, we will tell you so and suggest you get it from another evaluator.
Need to test an agent instead?
For evaluating an agent someone else built, or building your own evaluation system, see AI evaluation.
AI evaluation overviewStart with one workflow
A free 30-minute call to pick the workflow where an agent pays off fastest, and to decide cloud or local based on your data policy.