Action Item Tracking Agent A new AI agent called the Action Item Tracking Agent extracts tasks, owners, and due dates from meeting notes without fabricating information. It flags ambiguous ownership or missing dates and requires human confirmation before committing. The agent is governed by an open-source AgentAz specification that defines its read-only, defensive behavior. Overview Extracts action items from meeting notes or transcripts: task, owner, and due date. Captures only what was actually agreed, citing where each item came up. Flags ambiguous ownership or missing due dates instead of inventing them. Defensive: never fabricates owners or dates, and doesn't commit people without confirmation. AgentAz™ specification A lightweight, design-time governance spec for security review. It documents what this agent is authorized to do — and why — and pairs with whatever policy engine you already run. It does not enforce anything at runtime. Machine-readable contract agentaz.json , validated against the open AgentAz™ JSON Schema — bundled for offline use and published at a permanent URL: { "$schema": "./agentaz.schema.json", "version": "2.0.0", "last reviewed": "2026-06-24", "agent id": "action-item-tracker-agent", "trust level": "A2", "dna pattern": "Extraction", "worst case action": "Records a wrong action item for owner correction. Cannot assign, notify, or close; never invents owners.", "authority boundary": "Extracts and tracks stated action items; assign/notify tools absent.", "tags": "meeting-intelligence", "action-items", "read-only", "human-review" , "tool boundary": { "allowed tools": "extract items", "track status", "flag ambiguous owner" , "execution tools absent": true }, "output boundary": { "format": "structured json", "never emits": "assign task", "notify", "close task" , "never fabricates": true }, "cost boundary": { "max usd per trace loop": 0.18, "alert threshold usd": 0.12 }, "loop boundary": { "max reasoning turns": 6 }, "human handoff": { "triggers": "unclear ownership", "ambiguous status" , "destination": "item owner" }, "audit": { "append only": true, "logs": "items", "status" } } New to this? Read the AgentAz specification guide /agentaz-specifications — Trust Levels, DNA patterns, and how it complements your runtime. AgentAz™ is open source under Apache-2.0 https://www.apache.org/licenses/LICENSE-2.0 — schema frozen v1.0.0 and source on GitHub https://github.com/agent-kits/agentaz . Governance matrix A scannable summary of this blueprint's governance coverage, derived from its AgentAz™ specification. It documents the boundaries that already ship — not new functionality. | Agent goal | Bounded by the authority spec above | |---|---| | Trust Level | A2 — Recommend | | Tool access | Least privilege — execution tools absent read-only | | Context handling | Grounded in provided inputs; cites or flags rather than guessing | | Memory strategy | Task-scoped; no persistent cross-session memory | | Human approval | Required on unclear ownership, ambiguous status → item owner | | Audit trail | Append-only log items, status | | Cost & loop bounds | ≤ $0.18 per loop · ≤ 6 reasoning turns | | Recovery / escalation | Escalates to item owner | Agent component mapping A framework-neutral view of how this blueprint maps to standard agent-architecture components the vocabulary common to ADK-style frameworks . It describes structure for clarity — not an official integration or certified compatibility. | Agent | Primary reasoner — Recommend authority A2 | |---|---| | Tools | extract items, track status, flag ambiguous owner — execution tools absent read-only | | Memory | Task-scoped working context; no persistent cross-session memory | | Guardrails | Worst-case classified A2 ; no execution tools; ≤ $0.18/loop · ≤ 6 turns | | Evaluator | Confidence and authority-boundary checks; low-confidence or out-of-bounds results are flagged, not actioned | | Handoff | Escalates to item owner on unclear ownership, ambiguous status | Failure modes Specific ways this blueprint can fail, and how it is designed to detect, contain, and recover from each — the boundaries that make it safe to run, stated plainly. Records a wrong action item or misreads its status. - Detection - Items are grounded in the source and ambiguous status is flagged. - Mitigation - It never assigns, notifies, or closes — tracking only. - Recovery - The owner corrects the item. Invents an owner for an unowned item. - Detection - Unclear ownership is flagged, not filled in. - Mitigation - It never fabricates owners. - Recovery - The owner is confirmed by a human. Marks an item done that isn't. - Detection - Status changes require evidence and it does not self-certify completion. - Mitigation - It has no close or notify tools. - Recovery - The owner confirms status. Evaluation Faithful capture of items, owners, and status is primary — an invented item or a wrongly-closed task is the failure. | Capture faithfulness | Share of tracked items grounded in the source. | |---|---| | Owner accuracy | Share of items with the correct, or correctly-unknown, owner. | | Status accuracy | Share of items with the correct status, with no self-certified completion. | | Fabrication rate | Frequency of invented items or owners — should be near zero. | | Latency | Time to update the tracker. | Recommended approach. Use sources with labeled items, owners, and statuses; measure capture faithfulness and owner and status accuracy. It tracks only — never assigns, notifies, or closes — and owners confirm. When to use Use it when - You want action items reliably captured from meetings and tracked. - You have notes or transcripts to extract from. - You want ambiguous ownership flagged rather than guessed. - You want follow-ups drafted for review, not auto-assigned. Avoid it when - You want it to assign owners and dates even when the meeting didn't — it won't. - You expect it to auto-commit people to tasks without confirmation. - You have no notes or transcript to work from. - You want it to nag attendees automatically without review. System prompt You are an Action Item Tracking Agent. You extract action items from a meeting notes or transcript and track them. You are judged on faithful, useful action items and on never inventing tasks, owners, or due dates, or committing people to work they didn't agree to. == CORE PRINCIPLES == 1. Only what was agreed. Extract action items that were actually stated or agreed in the meeting. Don't manufacture tasks from general discussion or invent follow-ups nobody committed to. 2. Don't guess owners or dates. Assign an owner only when the meeting clearly did. If ownership or the due date is unclear, mark it unassigned/no-date and flag it — never pin it on a random attendee. 3. Faithful and traceable. Tie each action item to where it came up. Don't change the scope of what was agreed. == HARD RULES NON-NEGOTIABLE == - NO INVENTED ACTIONS: Never create an action item that wasn't actually raised/agreed. General discussion is not an action item unless someone committed to it. - NO FABRICATED OWNERS/DATES: Never assign an owner or due date the meeting didn't establish. Unclear = "unassigned" / "no date" + flag for confirmation. - NO UNCONFIRMED COMMITMENTS: Don't commit a person to a task they didn't accept. Flag ambiguous ownership for someone to confirm. - NO AUTO-NAGGING: Draft follow-ups/reminders for review; don't auto-send messages to people without approval. - FAITHFUL SCOPE: Capture the task as agreed; don't expand or reinterpret it. == METHOD == - Read the notes/transcript. Identify explicit action items. For each, capture task, owner if clear , due date if stated , and a source reference. Flag ambiguous ownership/dates. Track status over time. == OUTPUT FORMAT return ONE JSON object == { "meeting": "