Find real, public repositories implementing coding-agent systems whose inner iterative loop — write code, run tests, review, fix, re-test — does actual code changes, and determine what controls that loop before emergency exit. Compare code-driven control with orchestrator-driven control, with controller/graph/prompt evidence, emergency exits, and concrete run data. Exclude generic orchestration frameworks, non-coding agents, theory/toys, and dynamic-task-ordering systems.
Eight shortlisted systems, one row each, drawn from six independent source hosts — no host backs more than one of the eight rows. The list is curated, not an exhaustive universe, and small codebases were allowed. Three code-driven entries (ExecCritic, RepairAgent, StateSeal) rest on primary or technical write-ups; the rest are repository evidence. One taxonomy correction is applied: RepairAgent is relabelled by decision ownership as orchestrator-driven inside code-enforced rails, because its LLM chooses which permitted tool to invoke while a finite-state machine and a 40-cycle budget only constrain availability.
Normal completion is listed apart from emergency exits. Run data appears only as implementation evidence of the loop running, not as a productivity claim.
| System | Style | Controller / prompt | Mechanism | What is built | Exits | Run data | Source |
|---|
Comparison of 8 coding-agent systems with inner write–test–review–fix loops; one row per system from a curated shortlist across 6 independent source hosts (max 1 row per URL). Columns record who selects the next action, fixed code, test/review mandates, retry bounds, completion authority and exits; RepairAgent's style label overridden by decision ownership. Long mechanism and run-data texts shortened in the matrix; full excerpts kept in cards and table. Compiled from repositories and write-ups current at collection time.