AgentMES · manufacturing execution for autonomous agents

⚙ AI-Native Engineering Workflow

OpenAI's 7-stage bipartite pipeline — agents do the throughput work, humans own the judgment gates. Two tickets flow in parallel through Plan → Design → Build → Test → Review → Document → Deploy.

Blaxel × Codex × Redis AI Hack Day Wordware HQ — San Francisco 2026-04-11 github.com/benikigai/agent-mes

↑ space play/pause · ← → seek ±5s · . frame step · auto-pauses at each stage — press space to advance

STAGE 1

Plan

Natural-language compiler turns a Slack rant into a typed acceptance-criteria JSON.

Opus 4.6 classifies intent → structured AWP payload

STAGE 2

Design

Hydrate memory_provenance and context_bundle from long-term + schema-typed sources in parallel.

Agent Memory recall + MCP-typed ground truth

STAGE 3

Build

Asciinema replay of a real Codex worktree run — the diff is assembled stroke by stroke.

+47 / −3 lines across 3 files

STAGE 4

Test

Hardware-isolated microVM runs the suite inside the blast-radius the plan declared.

iter1 FAIL → iter2 egress KILL @ 23ms → iter3 PASS

STAGE 5

Review

Cross-check every memory claim against ground truth. Memory drift catches trigger a human gate.

⚠ DRIFT → ✓ HUMAN approved

STAGE 6

Document

Write the decision log back as a long-term lesson — negative constraints flagged for future recall.

lesson written · negative_constraint=true

STAGE 7

Deploy

Open a real PR with the full receipt trail as the body. Monitoring breadcrumb logged for standby.

PR opened · standby=blaxel · merged ✓

What you're watching

Two tickets flow in parallel through the same 7-stage pipeline. Cards on the left start as raw Slack text; as they progress they accumulate stage receipts inside the card body — by the time a card lands in DEPLOY, the body is the audit trail. The replay auto-pauses at each stage so you can read the per-stage breakdown above before continuing.

Ticket	Type	What it tests	The wow beat
⚙ TKT-001	CODE	Fix the flaky `test_oauth_token_refresh` blocking CI	Stage 4 Blaxel egress kill at 23ms · Stage 5 Redis drift catch on the prior mock
✉ TKT-002	SIMPLE	Draft postmortem for incident-2026-04-09	Stage 5 Redis déjà-vu catch — same root cause as a prior incident, action item never deployed

Run it locally

cd ~/code/blaxel-codex-redis-hackathon
source .venv/bin/activate
agent-mes demo --dry-run      # terminal kanban
agent-mes web                 # interactive web kanban on :8000

The terminal kanban needs ≥180×50 to render the card bodies without aggressive line wrapping.