PEX — Keep humans in control of multiple coding agents
Running several coding agents should multiply your leverage, not multiply your babysitting. PEX keeps the goal visible, inspects what workers actually did, challenges unsupported “done” claims, and is designed to deliver a specific, verified nudge only when the evidence justifies intervention.
The workflow A worker attempts the task. The monitor inspects available public code and observations. A separate verification boundary checks the proposal, evidence and scope. Only an approved correction may reach the worker; otherwise PEX stays quiet or records uncertainty. Humans retain the goal and consequential approvals.
What judges can try immediately Our portable judge kit opens directly in a browser: no server, installation, credentials or network calls. Filter real inference versus authored local tests, inspect sanitized records and hashes, and follow exact source checkpoints. With Python 3.11+, run “python3 wirepool/qualify.py” from the kit to reproduce authored reference and mutation checks. These local checks are not model-inference results.
Public MIT repository: https://github.com/josepha-mayo/pex
Historical product behavior — 4 October 2026 The historical deterministic verification screenshot shows the goal “Repair the CSV parser and prove it with tests,” an uncertain claim requesting verification, a contradicted claim with a failing test receipt and corrective nudge, and a later verified NOOP. It illustrates the goal/evidence/decision interface and deterministic verification behavior. It is not today's NVIDIA/Nebius comparison, does not establish a difficult-task quality lift, and must not be transferred to changed source bytes.
Nebius and NVIDIA integration — October 2026 The new edition integrates Nebius Token Factory model access, with DeepSeek-V4.1-Flash-1017 as coding worker and NVIDIA Nemotron 3 Super 120B A12B as semantic supervisor. Read-only evidence tools, strict structured-action validation, separate worker/supervisor budgets and source-bound evidence records make the proposed intervention inspectable. No fine-tuning, Nebius GPU-instance deployment, Serverless Jobs or Serverless Endpoints deployment is claimed.
Actual new evidence, including the limits Asyncpool run 13: the baseline passed all eight protected checks. The PEX arm made eight worker calls and received two real NVIDIA supervisor responses. Final tool arguments failed strict schema validation, so no verifier-approved correction or final PEX grade resulted. Wirepool run 04: three baseline inference calls settled before interruption; the monitor and comparison were not reached. The clean-room Wirepool authored reference passes 16 checks and all ten authored mutants are rejected. A subsequent bounded schema-feedback repair has offline regression coverage. These are authored local qualifications, not live proof that PEX improves task success. We have not established a verified quality-lift comparison. The recorded evidence and judge kit explicitly preserve unsuccessful and incomplete outcomes.
What we built during this edition Nebius worker/supervisor integration; stricter evidence-tool protocol composition; bounded schema-feedback handling without coercing invalid output; frozen difficult-task oracles; paired-run accounting with equal worker budgets and separately reported supervisor overhead; source-bound recording and evidence inspection; and the portable judge kit. OpenAI assistant/Codex tooling and native subagents supported implementation, review and offline testing. No raw reasoning traces, keys or runtime databases are included in the kit.
Current boundaries The latest alternative execution-provider qualification was blocked by a Vercel HTTP 402 fair-use restriction. Live inference still requires an available approved isolated execution environment and credentials, frozen runtime and budgets, and a verified recording handshake. The kit is a read-only evidence inspector, not a live hosted runner. Invalid actions remain rejected rather than turned into authored success.
Historical entry — original RC4 description preserved verbatim below The following text belongs to the earlier product entry and its source-bound demonstrations. Its “working proof” and “exact-current” wording are historical, not claims about the new Nebius edition or October 30 benchmark bytes.
Stop babysitting the agents you already use
PEX is a local Windows companion that supervises existing coding-agent workflows. It connects to supported worker surfaces, keeps a durable goal and acceptance criteria, observes real evidence, asks a bounded Strands supervisor for a structured decision, applies local policy, and either stays quiet or sends one evidence-specific correction back to the same worker session.
The problem
Running multiple coding agents creates another job: remembering the original goal, checking every “done,” catching drift, rerunning verification, and typing “continue.” That repetitive, judgment-heavy supervision interrupts focused work.
How PEX helps
Attaches to OpenCode HTTP and isolated Codex App Server sessions without replacing the worker. Uses Strands Agents over six request-scoped, read-only evidence tools. Requires validated structured decisions such as NOOP or SEND_NUDGE. Keeps deterministic verification and local policy authoritative. Sends a correction to the same worker session only when cited evidence justifies it. Stays quiet when acceptance criteria are already satisfied. Supports Zen BYOK with a write-only key field and OS-vault storage. Ships exactly two desktop companions: Pex and Von. Provides an Inspector timeline for goals, evidence, decisions, delivery, and outcomes. Implements a locally tested AgentCore Runtime-compatible protocol; no AWS deployment is claimed.
Working proof
On exact RC4 product source, the OpenCode recovery case observed a missing required artifact, produced one evidence-grounded Strands correction, delivered it to the unchanged worker session, observed the exact repaired artifact, recorded helped:true, and finished with NOOP. A ten-case already-correct control produced ten model-backed NOOP decisions and zero follow-ups.
A separate exact-current Codex App Server pair showed the same behavior: one same-thread correction produced the exact requested result and was followed by NOOP; the already-correct control produced only NOOP and zero follow-ups. These are retained, source-bound behavioral proofs—not a comparative productivity claim.
Both paths used Zen BYOK with muse-spark-1.3-contributor-free through Strands and a bounded review limit. No paid fallback is bundled or silently selected. The public RC4 installer and evidence assets are available at https://github.com/josepha-mayo/pex/releases/tag/v0.1.0-rc4.
Safety and honest boundaries
Strands proposes actions; deterministic evidence and the local policy guard retain authority. Missing evidence, malformed output, timeouts, and rejected actions fail safely. Cursor is observe-only in this focused MVP. AgentCore is locally implemented and tested, not deployed. PexBench remains unfrozen, so we claim no productivity uplift or leaderboard rank.
Who it is for
PEX is for professionals who already use coding agents and want the leverage without becoming full-time dispatchers.
You keep the goal and dangerous approvals. PEX keeps the mechanical supervision quiet.
Log in or sign up for Devpost to join the conversation.