AgentOps 3.0 Explainer Kit

This is the public text kit for explaining AgentOps 3.0 in a gist, README section, launch post, or video description.

One-Sentence Positioning

AgentOps is the engineering operating system for agent teams: a disciplined engineering layer that gives coding agents shared domain context, review verdicts, tracked follow-up work, and optional out-of-session compounding (the loop dispatched on a reference NTM + MCP + managed-agents substrate).

Problem Statement

Coding agents are useful, but most teams still operate them like isolated chat sessions. A model can make a reasonable product, design, or engineering call in one window, while the next run loses the domain language, prior concerns, review evidence, and follow-up decisions.

Human engineering teams solved this class of problem with shared domain models, specs, tests, code review, issue trackers, release gates, runbooks, and wikis. AgentOps encodes that operating discipline for agents.

The packet is the linked intent object. It is the thing that moves through the AgentOps lifecycle: product/domain intent becomes a context briefing, then a council verdict, then tracked work, then an execution packet, validation evidence, and a handoff. Provenance and trace make that movement inspectable.

Canonical doctrine for that discipline lives in docs/cdlc.md. PRACTICE-REGISTRY.md backs the packet with practice lineage and stable practices: [slug] citations.

Target User

AgentOps 3.0 is for agent-heavy maintainers, staff engineers, technical founders, and small teams who already use Claude Code, Codex CLI, Cursor, or OpenCode on real repositories and want repeated agent work to become more coherent instead of resetting every session.

They do not need a hosted control plane to see value. The first proof is one decision becoming a shared packet, a council verdict, and tracked work.

The 5-Command Path

After installing AgentOps and restarting the agent runtime:

ao quick-start
cp docs/examples/agentops-3-domain-practice-packet.md .agents/packets/agentops-3-launch.md
ao context assemble --phase planning --task "Evaluate the AgentOps 3.0 launch demo against the domain/practice packet" --output-file .agents/rpi/briefing-current.md

/council --mixed validate "Given .agents/packets/agentops-3-launch.md, should the AgentOps 3.0 launch demo lead with council-first engineering judgment?"

BEADS_DIR="$(ao beads dir)" br create "Apply council verdict to launch demo" --body "From .agents/council/<run-id>/verdict.md"

Expected artifacts:

• .agents/packets/agentops-3-launch.md
• .agents/rpi/briefing-current.md
• .agents/council/<run-id>/verdict.md
• _beads/issues.jsonl

For the full first-session path with time budget and fallbacks, see AgentOps 3.0 First-Value Path.

For launch video outlines, clip hooks, CTAs, and measurement fields, see AgentOps 3.0 YouTube Starter Series. For observed first-run evidence and claim promotion rules, see AgentOps 3.0 PMF Evidence Loop.

Domain Packet Example

Use AgentOps 3.0 Domain/Practice Packet as the launch example.

The packet makes these things visible before the agents judge:

• Product identity and target user.
• The decision under review.
• Product, goal, issue, standards, and evidence sources.
• The practice lineage and citations from PRACTICE-REGISTRY.md.
• Engineering practices to enforce, including DDD/TDD/BDD/review/release discipline where relevant.
• Non-goals and claims that require external evidence before public use.

Council Verdict Example

Use AgentOps 3.0 Council Verdict Example as public sample output.

The important shape:

• Every judge sees the same domain/practice packet.
• Claude and Codex can disagree, but they are disagreeing inside the same engineering frame.
• The consolidated verdict is PASS/WARN/BLOCK, not loose advice.
• The verdict creates follow-up work or a launch decision.

Why This Is Not Just Multi-Chat Prompting

Ad hoc multi-chat	AgentOps 3.0 path
Paste context into several chats manually.	Put the domain/practice packet in a reviewable artifact.
Ask each model for an opinion.	Ask judges for a verdict against the same evidence bar.
Copy useful parts back by hand.	Record .agents/council/<run-id>/verdict.md and create br follow-up work.
Lose the reasoning after the session.	Keep local packets, briefings, verdicts, issues, and learnings inspectable.
Re-explain intent at every phase.	Hand the packet lineage through briefing, verdict, execution, validation, and handoff artifacts.
Automation starts as a giant promise.	Automation is second-stage, after the packet and verdict earn trust.

Out-Of-Session Expansion Path

Out-of-session orchestration is the deeper lane, not the first proof. AgentOps ships no daemon or scheduler of its own — the loop runs in session, and unattended runs are delegated to a swappable orchestration substrate. The reference substrate is the trio NTM + MCP + managed-agents (see ADR-0009 and docs/dependencies.md).

After the user sees one packet and one verdict:

# The substrate dispatches a whole loop as one unit: it spawns an agent that
# runs the /rpi skill over the next ready bead — it never drives the loop's insides:
ao rpi <bead-id>
# Scheduled maintenance (Dream reports, wiki curation, release checks) is driven
# by the substrate's triggers (NTM swarm, cron, or MCP). See docs/dependencies.md.

Use the substrate lane for approved recurring work such as Dream reports, wiki curation, release checks, or other compounding jobs where the operator has already accepted the artifact shape.

Evidence-Gated Claims

Use these claims now:

• AgentOps is the engineering operating system for agent teams.
• AgentOps is a disciplined engineering layer for agentic software development.
• AgentOps turns agent opinions into engineering verdicts.
• AgentOps keeps agent work local, inspectable, and repo-native unless the operator chooses external services.

Do not claim these without exported evidence:

• Product-market fit.
• Specific productivity or speedup numbers.
• Safety-critical, regulated, or certified operation.
• Fully autonomous source mutation as the first-value promise.
• A public customer outcome that only exists in local .agents/ notes.

Launch CTA

Use one CTA at the end of public content:

Install AgentOps, run ao demo --quick, then follow docs/first-value-path.md to get your first council verdict.

The goal is not to explain every skill. The goal is to get a maintainer to see one shared engineering domain become one verdict artifact they can inspect.