Experience-Driven Black-Box Cost Control for Coding Agents
Python agentic-aiagentic-codingai-codingai-coding-assistantanthropicartificial-intelligenceclaude-codecodexcoding-agentcoding-agentscost-optimizationcost-reductiondeveloper-toolsgithub-copilotllmopen-sourceopenaiproductivitytoken-optimization
docs(experiments): point two links at te_bench/runner.py, where AgentRunner now lives
Aaron3b08fc17b210
diff --git a/experiments/docs/COPILOT.md b/experiments/docs/COPILOT.mdindex 5997a64..c7bbebc 100644--- a/experiments/docs/COPILOT.md+++ b/experiments/docs/COPILOT.md@@ -1,7 +1,7 @@# Running the GitHub Copilot CLI as the sealed agentTokenEconomizer's mechanisms are agent-agnostic, but each sealed agent needs its own runner-below the [`AgentRunner`](../../tokeneconomizer/runner.py) seam. This page is the Copilot path,+below the [`AgentRunner`](../te_bench/runner.py) seam. This page is the Copilot path,end to end, for **both** SWE-bench and Web-Bench. It assumes you have already done[`SETUP.md`](SETUP.md) steps 1–3 (venv, the two outside-the-repo directories, datasets);everything Copilot-specific is here.diff --git a/experiments/docs/SETUP.md b/experiments/docs/SETUP.mdindex 73da692..a5b22e5 100644--- a/experiments/docs/SETUP.md+++ b/experiments/docs/SETUP.md@@ -169,7 +169,7 @@ Copilot gate is in [`COPILOT.md`](COPILOT.md) Step 1.The heavy machinery — launching a sealed agent, applying the economizer's per-taskdecisions, grading against gold — sits **below `AgentRunner`**-([`../tokeneconomizer/runner.py`](../../tokeneconomizer/runner.py), which is the authoritative+([`te_bench/runner.py`](../te_bench/runner.py), which is the authoritativecontract). The repo ships a reference implementation, [`te_harness/`](../te_harness) — see[`HARNESS.md`](HARNESS.md) — so this step is a checkrather than something to write. Every launch selects a runner with `RUNNER=module:factory`