Agent infrastructure
What makes an agent work?
An agent harness is the software around a model that manages its tools, execution, context and task state. The model proposes actions; the surrounding system decides how those actions are carried out and recorded.
The model is one part of the system
The same model can behave differently with different tools, instructions, memory and execution environments. An SDK or framework may supply some of those parts. A hosted product packages them into an experience. Vendors use these terms differently, so compare the actual responsibilities rather than the label alone.
- Tools and execution
- Which actions are available, where code runs and how results return to the model.
- Context and state
- Which instructions, files, previous actions and checkpoints survive as work continues.
- Control and recovery
- Who authorises consequential actions, how a run stops, and what happens after a failure.
Long-running work needs continuity
Anthropic's published example uses an initial setup step followed by incremental work, progress records and checks. Manus describes file-based context and retaining useful error history. These are examples of system design around models, not guarantees that every task will succeed.
Compare the complete setup
- Record the model and harness versions used in a test.
- Check tools, permissions and where execution occurs.
- Inspect the trace and verify the actual deliverable.
- Measure the time, usage cost and human intervention for your workload.
- Test interruption, recovery and cases where approval is required.
A successful demo with broad permissions does not establish reliability under a narrower deployment policy. Keep the workload and environment comparable when interpreting a result.
Examples to explore
Manus is a hosted agent product. Deep Agents is a configurable harness. OpenHands provides software-agent components. Their profiles separate documented capabilities from tests we have not performed.