🚀 BMB v0.2 Released — Your AI Agents Now Learn From Their Own Mistakes

[project820]

TL;DR: BMB v0.2 adds a built-in analytics engine and a new Analyst agent. Your pipeline now tracks every agent spawn, timeout, and failure — then a dedicated agent reviews the data after each run and tells you what went wrong, what keeps repeating, and what to fix.

---

The Problem We Solved

In v0.1, BMB ran your multi-agent pipeline and delivered results. But if an agent kept timing out, or the same merge conflict happened three runs in a row, you had no way to know unless you were watching the terminal.

v0.2 fixes this with industrial-grade observability — borrowed from an unexpected source.

Bird’s Law: From Factory Floors to AI Pipelines

Frank Bird studied 1.7M industrial accidents and found a pattern:

For every 1 fatal accident, there are 10 serious injuries, 30 minor injuries, and 600 near-misses.

We applied the same ratio to AI agent failures:

Severity	Industrial	BMB Equivalent
600 (near-miss)	Close call	Routine file ops, normal lifecycle
30 (minor)	Property damage	Loop-back, timeout
10 (serious)	Injury	Agent crash, merge conflict
1 (fatal)	Fatality	System failure, full rollback

BMB now classifies every event using this model. Small repeated issues get flagged before they cascade into real problems.

What’s New

📊 Analytics Subsystem

Every pipeline run is now recorded in a local SQLite database. Agent spawns, timeouts, conflicts, loop-backs — all logged with 5W1H metadata. High-frequency identical events use count-based aggregation so your DB stays lean.

🔍 Analyst Agent (Step 10.5)

A new 9th agent runs after the Writer. It queries your analytics DB and produces:

• Session metrics and incident log
• Recurring pattern detection (across 3+ sessions)
• Agent reliability stats
• Timeout adequacy assessment
• Learning promotion candidates

It recommends — never auto-edits your config.

🎧 Consultant as Coordinator

The Consultant is now modeled after a heavy-industry site coordinator:

• Full situational awareness, zero command authority
• Dual-channel comms: real-time JSON events + durable narrative feed
• Post-briefing protocol: receives blind phase results only after verification
• Overtime nudging when agents run long

📋 3-Tier Reporting

Clear contract for what gets reported to you and when:

• Tier 1 (Immediate): Rollback, system failure, design change
• Tier 2 (Post-hoc): Library swap, agent respawn
• Tier 3 (Silent): Routine ops — you don’t need to know

📚 Context7 Integration

Executor, Frontend, and Architect now query live library docs via MCP. No more hallucinated API signatures.

By the Numbers

Metric	Value
New agents	8 → 9
Pipeline steps	11 → 11.5
Test assertions	121+
Verification rounds	4
New files	2
Modified files	6

Get Started

curl -fsSL https://raw.githubusercontent.com/project820/be-my-butler/main/install.sh | bash

Existing v0.1 users: just re-run install. Analytics activates automatically, all existing data preserved.

Links

• 📖 Full What's New
• 🗒️ Changelog
• 🌐 Interactive Docs
• ⭐ Star the repo if this is useful!

---

Questions, feedback, or ideas? Drop them in this thread — or open an issue.

[project820]

Interesting timing — @karpathy just posted about the same problem

Just hours ago (Mar 12, 2026), Andrej Karpathy posted a thread on X about the future of agentic programming that aligns remarkably with what BMB v0.2 is solving:

1. "We need a bigger IDE" — agents are the new unit of programming

"The basic unit of interest is not one file but one agent. It's still programming."

BMB treats agents as first-class units: 9 specialized agents, each with defined roles, timeouts, and lifecycle tracking. The pipeline is the IDE for orchestrating them.

2. "Agent Command Center" — real-time visibility into agent teams

"I feel a need to have a proper 'agent command center' IDE... I want to see/hide toggle them, see if any are idle, pop open related tools, stats (usage), etc."

This is exactly what BMB's Analytics subsystem does — every agent spawn, timeout, merge conflict, and loop-back is logged to SQLite with 5W1H metadata. The Analyst agent then queries this data and reports patterns.

3. "Human orgs are not legible" — but agentic orgs can be

"The CEO can't see/feel/zoom in on any activity in their company, with real time stats... with this level of legibility will that be optimal?"

BMB's answer: yes, full legibility is optimal. The Consultant Coordinator has complete situational awareness with zero command authority. The 3-tier reporting hierarchy filters noise so the human only sees what matters.

4. "Intelligence brownouts" — failover and resilience

"Have to think through failovers. Intelligence brownouts will be interesting."

BMB v0.2 addresses this with cross-shell state recovery (.bmb/analytics/state.env), the Analyst's timeout adequacy assessment, and Bird's Law severity classification that catches cascading failures early.

5. "Org code" — forkable agentic organizations

"You can't fork classical orgs (eg Microsoft) but you'll be able to fork agentic orgs."

BMB's recipe system (feature/bugfix/refactor/infra) is literally forkable "org code" — each recipe defines which agents participate and how they interact.

---

The alignment is almost uncanny. Karpathy is describing the problem space on Twitter; BMB v0.2 shipped working solutions for it today.

Would love to hear thoughts from anyone experimenting with multi-agent observability. What patterns are you seeing?