Harness Engineering
Harness engineering in 2026 boils down to four things:
- tool availability and use
- context management
- orchestration capabilities
- programmability
The frame I keep coming back to: the ideal agent platform is an intelligence layer running on top of a stateful block of data. The LLM supplies the intelligence; everything else the harness does is in service of feeding that intelligence the right context and letting it act. Once you hold it that way, the load-bearing design decisions fall out naturally.
What capabilities does the harness expose to the model? This is tool selection, and it’s the most obvious lever — but the subtle part isn’t which tools exist, it’s how discoverable and composable they are at the moment of decision. Too few and the model is blind; too many and the context budget (and the model’s attention) gets swamped. The good harnesses treat the toolset as something curated and situational rather than fixed.
How does the harness persist and expose session state? The “stateful block of data” is the real substrate. A harness is only as good as its conventions for reading, writing, and surfacing that state on top of the underlying storage layer. Context management is the visible face of this, but the deeper question is the access pattern: what state is durable, what is ambient, and how the model is expected to reach for it.
How easy is it to invoke programmatically — or recursively? Orchestration is where agents stop being chat and start being infrastructure. The interesting frontier is recursion: an agent tool that can spawn a fresh agent in the same workspace, hand it a task, and retrieve its output. Once invocation is cheap and composable, you get fan-out, delegation, and verification loops for free.
How easy is it to program the runtime prompt-space? This is the one that’s underrated. If the procedural layer — the prompt-space the runtime assembles each turn — is itself programmable, and you can do programmatic edits over the prompt layer, you open a meta-programming trajectory: the agent (or the operator) shaping the very instructions that shape the agent. That reflexivity is where the emergent behavior lives.