The code-as-tool-calling-layer section is the sleeper insight here. Moving from one tool call per LLM invocation to having the agent write a procedural loop once and execute it has cut our task completion times by roughly 3x on multi-step agent workflows.
The parallel trend of agents using filesystem context management to avoid context rot is where harness engineering is heading.
We dig into these patterns regularly in Agent Builders newsletter for daily fix on latest and cutting edge Agentic AI updates : https://aiagentbuilders.substack.com
The code-as-tool-calling-layer section is the sleeper insight here. Moving from one tool call per LLM invocation to having the agent write a procedural loop once and execute it has cut our task completion times by roughly 3x on multi-step agent workflows.
The parallel trend of agents using filesystem context management to avoid context rot is where harness engineering is heading.
We dig into these patterns regularly in Agent Builders newsletter for daily fix on latest and cutting edge Agentic AI updates : https://aiagentbuilders.substack.com
The code-as-tool-calling-layer point is spot on. Been building exactly this with OpenCode; a lead agent analyses a diff and routes chunks to specialist subagents in parallel. Each gets a focused context and a tight checklist. The orchestrator just synthesises. Covered the implementation here: https://reading.sh/one-reviewer-three-lenses-building-a-multi-agent-code-review-system-with-opencode-21ceb28dde10