An agent harness for long-running autonomous work
Claude-Code-style orchestration with planning, tool execution, retries and evals, so an LLM can carry multi-step tasks to completion without babysitting.
Sector: Developer tooling · details anonymized
The challenge
Single-shot LLM calls fall apart on real work. Multi-step tasks need planning, state, recovery from failed steps and a way to know when the job is actually done.
The goal was a harness, the machinery around the model, that turns a capable LLM into a dependable worker.
Our approach
We treated the harness as the product: task decomposition, a typed tool registry, checkpointed state so runs survive restarts, and automatic retries with backoff and strategy changes on repeated failure.
Every capability got an eval before it got a feature flag. Regression suites ran on each prompt and model change.
Architecture
A planner and executor loop with explicit state transitions. No hidden context accumulating in one giant prompt.
Sandboxed tool execution with per-tool permissions, timeouts and structured logging of every action the agent took.
Cost and token budgets enforced at the run level, with traces for debugging any decision after the fact.
Outcome
The harness pattern is now how we build all agentic delivery work at Mindela.
Building something similar?
We're happy to talk through how this architecture would map to your problem. No pitch, just engineering.
Start the conversationMore work
Multi-agent systems for autonomous driving at Uber
Consulting on how fleets of models coordinate for autonomous driving: agent communication, simulation-first evaluation and the verification gates between a model's decision and the road.
An AI assistant that resolves queries end-to-end
A tool-using AI assistant grounded in the client's own knowledge. It answers, acts, and escalates to humans when confidence drops.