Incremental Opportunities for Public Projects — 2026-08-11
Scope: this window's LongHorizon-Harness paper, the ECC agent-harness toolkit note, and LoopsBench, checked against Projects/public-projects.md's agent_harness project. The 2026-08-09 pass already flagged Living-Harness (self-evolving harnesses) as a fourth experiment candidate for agent_harness — this pass covers three additional, previously-unchecked sources landing the same week that suggest a different, complementary experiment rather than repeating the Living-Harness finding.
LongHorizon-Harness's manage-execute-audit split is a fifth candidate rung for agent_harness, distinct from Living-Harness's self-evolution rung
agent_harness already tests planning, memory, verification, sandboxing, and sub-agent delegation as separate reliability rungs. LongHorizon-Harness names a specific structural failure the current "verification" rung likely doesn't isolate cleanly: when the same context that executes a task also judges whether it succeeded, an incorrect self-assessment silently propagates into the next decision — the fix is splitting a manager (holds task state, decides what's next, never executes), a fresh-context executor (does the work, never judges itself), and a read-only auditor (verifies against the live environment, never acts) into three genuinely separate roles. The reported gains are large and reproduced across two different model families (Qwen 3.7-Plus and Claude Opus 4.7), which is a stronger signal than a single-model result. A concrete experiment: take agent_harness's existing verification rung and split it into this three-role structure specifically, then compare against the current single-context-verifies-itself baseline on the same benchmark tasks already in use — this tests a different, more specific claim than "does self-evolution help" (Living-Harness's question); it tests "does separating execution from self-grading help, independent of any learning/evolution mechanism."
The ECC toolkit's "review from fresh context" and AgentShield's pre-tool-use gating are both already-built reference implementations of rungs agent_harness currently only benchmarks in the abstract
This window's ECC note documents a shipped, MIT-licensed implementation of two patterns directly relevant to agent_harness's existing scope: a TDD workflow that explicitly spawns a fresh-context review step rather than letting the implementing context grade its own work (the same principle LongHorizon-Harness's auditor role formalizes above), and AgentShield/GateGuard — a deterministic pre-tool-use guard that blocks destructive shell commands before execution rather than relying on the model to self-police, which is a concrete instance of agent_harness's "sandboxing" rung. Since ECC is open-source and installable via /plugin marketplace add, it's a candidate to actually run against agent_harness's existing benchmark tasks as a fourth or fifth harness configuration, rather than only reasoning about the pattern abstractly — turning agent_harness from "we tested five rungs we built" into "we tested our own rungs against a shipped, community-maintained implementation of overlapping patterns."
LoopsBench's 25%-success-rate finding is a sobering calibration point for agent_harness's own benchmark scope
LoopsBench reports that even the strongest tested configuration (Opus-4.7 in Claude Code with an "outer continuation" strategy) resolves only 25% of its dependency-DAG-structured, long-horizon coding tasks — a result explicitly framed as evidence that most coding-agent benchmarks measure single-shot task success, not sustained multi-step reasoning across dependent units of work. If agent_harness's existing benchmark tasks are single-shot or short-horizon (worth checking against the project's actual task design), this is a concrete argument for adding at least one long-horizon, dependency-chained task to the benchmark suite — the rung-by-rung comparisons agent_harness already does would be more informative if at least one task category matched the shape LoopsBench argues most agent evals currently miss entirely.