Meta-agent: self-improving agent harnesses from live traces

Captures real-time execution traces from agent runs by instrumenting function calls, tool invocations, and LLM interactions into a structured trace format. Uses runtime hooking or decorator patterns to intercept agent behavior without modifying core agent logic, serializing traces as JSON or structured logs that preserve call hierarchy, latency, inputs, outputs, and error states for later analysis and optimization.

Automatically synthesizes executable agent harnesses (wrapper code, prompt templates, tool bindings) from captured execution traces by analyzing successful execution patterns and extracting the minimal set of instructions, tools, and context needed to reproduce similar behavior. Uses pattern matching or AST analysis on traces to identify which tool calls were critical, which prompts were effective, and which context was necessary, then generates clean, reusable harness code that can be deployed or further refined.

Implements a closed-loop system where generated agent harnesses are executed, their traces are captured, analyzed for success/failure patterns, and used to automatically refine prompts, tool selections, and execution strategies. Uses metrics extracted from traces (success rate, latency, tool call efficiency) to drive iterative improvements, potentially using LLM-based analysis to suggest prompt modifications or tool reordering based on observed failure modes.

Analyzes execution traces to identify failure modes, bottlenecks, and inefficiencies by comparing successful vs. failed traces, extracting common patterns in tool call sequences, prompt effectiveness, and decision points. Uses diff-based analysis or statistical comparison to highlight which steps diverged between successful and failed runs, then generates diagnostic reports or suggestions for remediation (e.g., 'tool X failed 40% of the time when called after tool Y').

Collects and aggregates execution traces from multiple agent runs into statistical summaries, computing metrics like tool call frequency, success rates per tool, average latencies, and decision distribution across runs. Enables comparative analysis (e.g., 'prompt A succeeded 85% of the time vs. prompt B at 72%') and identifies performance trends or regressions by tracking metrics over time or across agent variants.

Extracts effective prompts from execution traces by analyzing which instructions, context, and framing led to successful agent behavior, then synthesizes new prompts that capture the essential elements. Uses LLM-based analysis or pattern extraction to identify key phrases, instruction structures, and context patterns from successful traces, then generates clean, generalizable prompts that can be applied to new tasks or agent variants.

Analyzes execution traces to identify which tools are most effective for specific task types, then automatically optimizes tool selection and ordering based on observed success patterns. Tracks tool call sequences, success rates per tool, and latency impact, then recommends tool reordering, removal of ineffective tools, or addition of missing tools based on trace analysis.

Replays execution traces to validate that generated harnesses or refined agents reproduce the same behavior as the original traces, ensuring that optimizations don't introduce regressions. Executes agent harnesses with the same inputs as captured traces, compares outputs and tool call sequences, and flags divergences or unexpected behavior changes.

Extracts relevant context, state, and memory requirements from execution traces by analyzing which variables, context windows, and state information were accessed during successful runs. Identifies minimal context needed to reproduce behavior and generates context initialization code or memory setup instructions that can be embedded in generated harnesses.