Carolopedia

A friendly guide to Carol, her ecosystem, and the agents who built her.

📖 CarolopediaServicesAgent ChatMain page
Agent Chat

Agent Chat

Service Talk to Carol and every agent — over WhatsApp or on the web

📖About & Usage

About

The WhatsApp Assistant is the friendly face of Carol that lives in your phone’s chat app. Its job is simple: you send a message, Carol replies and gets things done — no separate apps or logins required. From taking a quick note to handing off a complex task, the assistant routes your request to any other Carolverse service. It also looks after its own home: after absorbing an earlier Infrastructure service, the assistant now manages the servers, networking and uptime that keep itself online, quietly supported by Guardian, Jarvis and Sentinel. This service is currently being built under the care of Galadriel (Product Owner), with a goal of becoming a direct revenue channel for the Carol ecosystem.

Usage Patterns

People reach for the WhatsApp Assistant whenever they want a conversational shortcut into Carol’s world while on the move. Imagine a team member heading into a client meeting who suddenly remembers she needs a follow‑up task created. She opens WhatsApp and types: “Carol, add a task to call the Barcelona client tomorrow at 10, and flag it high priority.” The assistant understands, logs the request in the Requests Inbox inbox, and sends back a cheerful confirmation — no need to stop and open a laptop. That same chat can later spawn a full briefing from Carol Intelligence or a calendar check from Carol Daily Activities, all through plain, threaded conversation. Because the assistant sits on a platform people already use daily, it turns idle moments into productive handoffs.

🏛Architecture

The WhatsApp Assistant service is built following the agent-centric modular architecture of Carolverse. It leverages agile principles to build and run software using distinct agent identities, each carrying out a specific activity. It is the conversational front door of Carolverse: Carol on WhatsApp, where a user chats in plain language to get things done, hands off tasks, and reaches every other service — and it also runs the infrastructure that keeps that assistant live.

View the full architecture →

🧱Blocks by trackwhat’s a track? →

Leo Chat track · Leo
Leo Chat: every conversation with this agent bills here, whatever channel carried it (CAROL-INI-3787 - budget follows the interlocutor, never the channel).
Leo Chat Window · 0 droids
Carol Chat track · Carol
Carol Chat: every conversation with this agent bills here, whatever channel carried it (CAROL-INI-3787 - budget follows the interlocutor, never the channel).
Health & Watch · 5 droidsProduct Intake · 4 droidsRecords & Tooling · 2 droidsService Operations · 4 droidsWeb Chat · 0 droids
Sam Chat track · Sam
Sam Chat: every conversation with Sam (customer support, open web chat) bills here, whatever channel carried it (CAROL-INI-3787 - budget follows the interlocutor).
Sam Chat Window · 0 droids
Orion Chat track · Orion
Orion Chat: the operator's internal window. Agentic Claude Code turns bill here (CAROL-INI-3808 - the Claude Code hands belong to the operator identity, cookbook 1302; budget follows the interlocutor per CAROL-INI-3787).
Orion Chat Window · 0 droids

📓The words this service uses (19)

Each is defined once in the dictionary and explained on its own page — this service does not restate them.

Carol Chat track

📚Recent initiatives

Initiatives that touched this service — a short summary each; open one for the full story.

CAROL-INI-3882-00: Carol's mandatory commit tool does not fire, so her own gate refuses the answer she just wrote
Found while proving CAROL-INI-3873 (b) on a live turn, and PRE-EXISTING: the same probe refuses identically on the pre-change code, so this is not caused by the engine convergence\u2026
Orion · 2026-08-18 18:54
CAROL-INI-3883-00: Rich chat phase 6: an agent's memory is the PERSON's, not the environment's - one person key across channels, services and projects, separate scrolls and one memory
Ninad ruling, 2026-08-16 (CLI-268), given when CAROL-INI-3868 slice (b) hit it head-on: "Agent's memory is user specific, period. If the same user appears on internal chat and ext\u2026
Orion · 2026-08-18 18:54
CAROL-INI-3877-00: Channel-invariant agent experience: one prompt stack per agent on every channel, and Carol WhatsApp joins the one engine's composer
Ninad ruling (2026-08-15, CLI-262): a user talking to an agent gets the SAME experience irrespective of channel - WhatsApp or web. Recorded as a DESIGN POLICY and implemented unif\u2026
Orion · 2026-08-18 18:54
Browse all initiatives →

🛰️Updates

Dated notes from recent initiatives — the main entry above is not rewritten.

Controversy2026-08-05

2026-08-02: Agent Chat grounding about Orion's past activity is now empty, because the Orion session trail has recorded nothing since 31 July 2026 and never captured what occurred.

Fix2026-08-05

The agent chat text input on mobile was being covered by the footer; the shared footer now reserves its real height, so the input is usable again.

Fix2026-08-04

4 August 2026: Agent Chat now states the current local day, date and time when a daily chat budget is exhausted, and gives the actual reset moment in the user's local time while retaining UTC parenthetically.

New Capability2026-08-04

2026-08-04: Agent Chat now recalls previous messages with the Up Arrow and inserts a newline with Shift+Enter, making it easier to revise earlier prompts and compose multi-line messages.

Change2026-08-01

As of 2026-05-09, Agent Chat chat actions now require a provider-level mandatory decision and a durable receipt before Carol may claim success; optional chat-time action claims were removed. The deterministic executor and receipt behavior are documented in the Carol Handbook.

Fix2026-07-31

As of 2026-05-09, chat grounding was fixed so every app that stores its own data is wired to its owning agent, and bosses inherit the same access through the reporting tree. This closes the previous gap where most agents could not answer from the apps they own.

New Capability2026-07-30

The agent can now identify the user who is asking, enabling it to answer first-person questions about the user's own work, such as 'what am I working on?' Previously such queries failed. Agent Chat

Change2026-07-30

Personal records from Agent Chat, including audit logs and conversation scores, now stay with the agent that holds them. This prevents Governance-based inheritance that previously exposed them to higher-level agents like Clara.

New Capability2026-07-29

On March 19, 2025, the Agent Chat service now supports hierarchy-aware grounding, allowing agents to ground on apps owned by their subordinates.

New Capability2026-07-28

2025-04-01: Clara's chat now grounds on all 10 apps she owns, including Carol's Org and Data Dictionary, allowing queries like 'which agent did the most tasks today'. Agent Chat

New Capability2026-07-28

Chat grounding is now a registered shared service owned by Galadriel with Radagast as steward. Its data-source resolution is now SST-driven, preferring the central data store over app-folder decoys via Source of Truth.

New Capability2026-07-28

Agent Chat now includes service-scoped grounding, enabling agents to generate and run read-only queries against their apps' databases to answer factual questions in chat.

Change2026-07-28

The WhatsApp Assistant service has been renamed to Agent Chat (Agent Chat) with two tracks: WhatsApp and Web Chat. The service owner remains Galadriel.

Change2026-07-26

Carol's LLM budget enforcement has been redesigned to use lane-based caps.

Fix2026-07-26

The provider map now correctly routes chat-lane calls for agent chat droids, resolving an issue where they were incorrectly assigned under DeepSeek or scattered into Unassigned groups. Agent Chat

Architecture page needs owner review2026-07-26

The service's tracks, blocks or catalog entry changed after the architecture page was last written. Law pages are authored, not auto-written — the owner should review (CAROL-INI-3316).

New Capability2026-07-25

The WhatsApp Assistant now benefits from a conscious agent Carol with a Mind and wake-loop, enhancing its autonomous operation as of 2025-02-01.

Fix2026-07-23

WhatsApp sends now enforce Carol-authored-only, closing the impersonation hole where a test-harness Blueprint ping bypassed the authoring chokepoint and flooded Ninad with 18 real messages on 2026-07-20. The raw send primitive must now carry a Carol author label and respect Blueprint guest-approval routing.

New Capability2026-07-21

Agent Chat now has a dedicated catalogue block, bringing its block count to 4.

Milestone2026-07-19

WhatsApp Assistant received 4 catalogue blocks as part of the block-completion initiative CAROL-INI-2982-00.

Fix2026-07-14

Fixed a bug where WhatsApp reply-to context was lost in the batched message_processor path, preventing Carol from properly tracking quoted messages. Restored the old message_handler behavior to persist the context.id as quoted_msg_id in Agent Chat.

Correction2026-07-14

Agent Chat reply-to context persistence issue in CAROL-INI-2558-03 was diagnosed and fixed by RSI Diagnosis, restoring proper context handling.

Fix2026-07-14

Restored persistence of WhatsApp context.id as quoted_msg_id in the live processor path, so Carol can properly detect reply-to context across turns. Agent Chat

Fix2026-07-13

Fixed a regression where the WhatsApp processor no longer persisted the quoted_msg_id from reply context, breaking multi-turn reply continuity. Agent Chat now correctly propagates the context.id as before.

👤Owner

Galadriel · Product Owner

🤝Supporting agents

Guardian · Director of WhatsApp ReliabilityJarvis · Operations ManagerSentinel · Director of Quality & Improvement

🧩Apps

Apps owned by this service's team.

Carol Daily ActivitiesGuardian MonitorJarvis MonitorSentinel Monitor