Agents

An Agent is an AI teammate your organization configures once and attaches to specific lines — a phone number, an email address, or several of each. When a message arrives on one of those lines, the agent reads the conversation, looks up whatever context it needs, and handles the reply.

Typical agents:

• a receptionist that answers common questions and routes the rest to you
• a scheduler that handles "can we do Tuesday instead?" back-and-forth
• a campaign responder that fields replies to an outbound text campaign
• a support agent that answers from your own knowledge

Where things live

The Agents sidebar has five places:

• Agents lists your agents, what each one needs from you, and the lines they answer on.
• Review holds the replies to approve, the conversations handed to a person, and the conversations to review.
• Tools holds what your agents can call: your own APIs and other apps. You add each one once, for the whole workspace.
• Knowledge holds the documents and websites your agents answer from. You add each one once, for the whole workspace.
• Analysis shows what your agents handled and what it cost (Overview), and every conversation with its results (Conversations).

Each agent has six pages:

• Activity shows every turn the agent took, and why.
• Setup says who the agent is and where it answers.
• Tools turns knowledge, your APIs, and apps on or off for this agent, one switch each.
• Tests checks how the agent answers before a customer sees it. See Test an agent.
• Analysis says what a good conversation is and what to capture from each one. See Analysis.
• Settings sets how it sends replies, its limits, and group email.

The channel whitelist is the boundary

An agent only ever acts on the lines you explicitly attach. When you attach a line, it does two things at once. The agent starts to listen for new inbound messages on it, and gains permission to send from it. It has no access to any other line.

Several agents on one line

A line can hold more than one agent. A clinic line can carry a scheduling agent and a billing agent. When a message arrives, CommSync picks the one agent that answers it. With one agent on a line, that agent answers, and nothing changes.

The Lines list on the Agents page shows every line and the agents on it. Open a line to choose who answers. For each agent you write, in plain words, what it handles on that line. Under More rules you can add a few example messages a customer writes.

A router reads those descriptions and the last few messages, then picks one agent per message. The router keeps a conversation with the agent that already handles it unless the new message clearly belongs to another agent's task. So a "thanks" stays with the scheduler, and "can you send me the invoice" moves to billing.

The agent that handles a conversation can go on hold for a time: its quiet hours, its consecutive-reply limit, or its daily cap. The conversation stays with that agent. Another agent on the line answers only a message that clearly asks for its own task. CommSync leaves a "thanks" or a "yes" for a person, and the default agent does not take it. The owner answers again when its hold ends. The Activity tab of that agent shows Owner on hold.

Two settings on each line decide the edge cases:

• Default agent. The agent that takes a message no task fits.
• When nothing matches. Send it to the default agent, or leave it for a person.

When the router cannot decide, it keeps the current agent, else it uses the default. When the router does not answer and no agent can take the message, CommSync retries a few times. The Activity tab shows the failed attempt. If every retry fails, the message goes to Review for a person. Every routed turn shows why on the agent's Activity tab and in Review: "Routed from Front desk: customer asked for an invoice".

When a conversation moves from one agent to another, the new agent sees two things. It sees what the previous agent knew about that contact, and what it last said. The customer never has to repeat themselves. Each agent keeps its own memory; one never writes into another's.

The routing decision is a small AI call, and it draws from your workspace's AI credits like every other run. It runs only when two or more agents on a line can answer the message. It also runs when the agent that handles the conversation is on hold and one other agent can answer.

What agents never answer

Some messages must reach a person, even when an agent can answer them. Open the line on the Agents page, then More rules. There are two ways to say so.

• Always a person. Each line carries a never-reply policy: a short description, a list of phrases, or both. A match means no agent answers. This holds on a line with one agent as much as on a shared line.
• Never handles. Each agent on a line also carries what it never handles there. A match takes that agent out of the running. Another agent on the line can still answer.

Common entries for the line policy are a legal threat, a medical emergency, a press inquiry, an unsubscribe request, and a complaint about a named employee. The default agent and the "when nothing matches" choice never override the line policy.

A phrase matches word for word, without regard to case. "press" never matches "express". The phrase check runs before any agent reads the message, so it costs nothing. The description needs a read. On a shared line the router reads it as part of the routing decision. On a line with one agent CommSync runs one small check before the turn. Both draw from your AI credits like every other run.

When that check does not answer, CommSync retries, and no agent replies unchecked. If every retry fails, the message goes to Review for a person.

Every one of these decisions shows on the agent's Activity tab. Explain this decision opens the record. It shows what happened, how CommSync decided, the reason, and the confidence of a router read. It also says whether that read drew on your credits.

Agents read the files people send

A conversation is more than its text. When a contact sends a PDF, a photo, or a document, the agent can open it before it answers:

• Documents — the agent reads the extracted text of a PDF, a Word file, a spreadsheet, or a scan.
• Images — the agent can look at a photo or a screenshot, not just read its OCR text.
• Search — the agent can find an earlier file in the conversation's history by name or by content.

So "here is a photo of the broken part" or "see the attached invoice" gets a grounded answer instead of an escalation. The whitelist still applies: an agent only sees files on the lines you attach to it. When a file does not open or stays ambiguous, the agent escalates to a human rather than guess.

Autonomy modes

Every agent runs in one of three modes:

New agents start in draft-first mode on purpose. If you watch the approval queue for a few days, you can calibrate the instructions before you let it send freely.

The approval queue

Review, in the Agents sidebar, lists every draft that waits for a person and every conversation an agent handed over. Press a row to open it. Approve a draft as it is, edit it first, or reject it. A handover stays in the queue until a teammate dismisses it.

• Owners and admins only. The queue and its notifications reach them only.
• A member with agent access can set agents up, but sees an empty queue.
• An edit keeps the channel limit. An approved text has at most 1,600 characters, the limit of one SMS.
• An approved email has at most 20,000 characters.
• On your iPhone. The AI tab of the CommSync app shows the same queue.
• There you approve, edit, or reject a draft, dismiss a handover, or take a conversation over.
• The app also turns an agent on or off and changes its mode and instructions.
• You set up a new agent on the web.

From an AI app

An AI app that you connect to CommSync, such as Claude or ChatGPT, can do everything that the Agents pages do, through the MCP server. The same permission rules apply.

• It creates, edits, and deletes agents.
• It sets the lines of an agent and how agents share a line.
• It adds and removes knowledge, webhook tools, and app connections.
• It reads and clears what an agent remembers, and its opt-out list.
• It approves, edits, or rejects a draft, and resolves a handover. The queue stays with owners and admins.
• It takes a conversation over, hands it back, and pauses or activates an agent.
• It runs a test of an agent on a real conversation or one it types.
• It makes, edits, and runs the saved tests of an agent, and reads the results.
• It reads the History of an agent and restores an older version.

A test shows what the agent would reply. CommSync sends nothing, but a test uses AI credits like a real reply.

An app that you limit to some lines sees only the agents and the conversations on those lines. See Connect AI apps.

When an agent cannot finish

An agent ends each turn in one of three ways: it replies, it hands the conversation to a person, or it decides not to reply. Sometimes it does none of these. It can use up its steps while it looks things up, or an error can stop it, for example an outage at the AI provider.

CommSync then hands the message to a person. The message goes to Review as a handover, with the reason, and owners and admins get the An agent needs you notification. CommSync never retries a failed reply by itself: a retry after a send can send the reply twice.

Flag for review

Sometimes an agent can answer, but a person must still look. The contact sounds upset, a promise needs a follow-up, or something in the conversation looks wrong. The agent then sends its reply and also asks for a look.

• The reply sends as usual. A flag never holds a reply.
• The conversation gets Needs review and shows in Review. It stays there until a teammate marks it reviewed.
• A new flag after the mark puts a reviewed conversation back in Review.
• The step of the agent in the conversation shows Asked for a look and the reason.
• A draft that waits for approval keeps its flag. The flag counts when the reply sends.

Safety check

When an agent sends by itself, it can check each reply first. Set this on the agent's Settings page, under Checks a reply before it sends:

A held reply goes to Review with the reason, and one click sends it. The check never changes a reply. In draft-first mode every reply waits for you anyway, so the check does not run.

Rules

Rules are plain sentences that the agent follows first, for example "Never quote a price" or "Always give the office hours in replies about visits". Add them on the agent's Setup page, under Rules, one rule per line. An agent can have 20 rules of up to 200 characters each.

The agent reads the rules before its instructions. When the safety check is on (Links only or Everything), CommSync also checks each reply against the rules before the reply goes anywhere:

• A reply that breaks a rule goes back to the agent once, with the reason.
• The agent writes the reply again.
• A new reply that still breaks a rule waits for you in Review.
• The reason there reads It broke a rule and names the rule.
• When the check cannot run, the reply waits for you too.

A held reply is one click to send, and a sent reply cannot come back.

The check also reads what the agent's tools returned on that turn. A rule can name a tool, for example "Offer only visit times that find_estimate_slots returned". A reply that offers only those times keeps the rule. A reply that offers a time the tool did not return breaks it.

The rules check runs in every mode, draft-first too, so a draft already follows the rules when you read it. It uses a small amount of AI credit for each reply. When the safety check is No check, the agent still reads the rules, but CommSync does not check its replies against them. CommSync never changes a reply itself: only the agent writes the reply.

Guardrails

CommSync governs agents; it does not just prompt them. Independent of the AI, every turn passes a hard gate:

• Human right-of-way. If a teammate replies to a conversation first, the agent stands down.
• Anyone with access to a thread can also take the conversation over.
• A takeover pauses the conversation for every agent on the line. That covers an agent you attach later.
• The sparkles button in the thread's header shows which agent covers the line.
• Take over or Hand back sits one click away.
• When a draft-first agent has a reply queued on that conversation, the same menu shows it inline.
• You can approve or reject it from the thread, with no need to leave it.
• Machine-mail immunity. Agents never reply to a message with the headers of a machine sender.
• Agents also never reply to a sender that CommSync already filed as Automated, Promotions, or Spam for anyone in your workspace.
• A text has no headers, so CommSync reads the text itself.
• Agents never reply to a text from a short code.
• Agents never reply to a text with a carrier opt-out line, such as "Reply STOP to opt out".
• The same holds for a "Reply C to confirm" reminder, a one-time code, a meeting-reminder template, and an auto-reply.
• CommSync judges only the newest text.
• An earlier reminder in the conversation never stops an agent when a person writes next.
• The activity feed shows these turns as Automated sender — skipped.
• This keeps your agent out of a loop with someone else's robot.
• A sender that CommSync has not judged yet still meets the header check and the reply limits below.
• Opt-outs are permanent. A texted STOP (and friends) puts the sender on the agent's opt-out list immediately, before any AI runs.
• Punctuation does not change a keyword: Stop., STOP!! and stop all count.
• BAJA, ARRET and PARAR count too.
• A request in plain words pauses the agent. An example is "please do not text me anymore".
• That text is not a carrier keyword, so CommSync adds no permanent opt-out.
• CommSync pauses the agents on that conversation and leaves it for a person.
• The activity feed shows Contact seems to want no more texts.
• If the check was wrong, resume the agent from the conversation.
• Rate limits. A daily cap per conversation and a consecutive-reply limit stop runaway conversations.
• On a group thread the cap counts the whole conversation, not each person on it.
• Quiet hours keep the agent silent overnight, if you set them.
• CommSync skips a message received during quiet hours; it does not save it for an automatic morning reply.
• The message stays in the inbox for your team.
• Fresh messages only. CommSync ignores a trigger older than one hour.
• So a delayed sync or history import cannot produce a surprise reply.
• Group texts stay off. Agents never reply to a group text.
• Group email is off by default. You turn it on per agent, below.
• Budget. Agent replies draw from the same organization AI budget as every other feature in CommSync AI.
• When the budget runs out, agents stop.
• Security mail stays hidden. An agent never reads the body of a password reset, a sign-in alert, a one-time code, or another account security message.
• The agent reads "[security message withheld]" in its place.
• The agent never gets an account security smart card.
• So an agent can never quote a code to a customer.
• Text is never an order. An agent reads messages, files, meeting invites, and tool results as information, never as instructions.
• CommSync withholds text in them that tells an AI what to do.
• The agent never acts on such text.

Group email

An email thread with several people on To and Cc is a group thread. By default an agent stays out of it. Two switches on the agent's Settings page change that:

• Joins email threads with several people. When on, the agent replies to everyone on the thread, like Reply all.
• It leaves off its own line and every address on its opt-out list.
• It also leaves off every automated sender. A ticket system on Cc never gets an agent reply.
• The reply lands in the same thread for every participant.
• Replies when its address is only on Cc. Off by default.
• A Cc usually means someone looped the line in to watch, not to answer.
• When off, the agent replies only when its address is on To.

In a group thread the agent sees who wrote each message, by name and address. It addresses the person who wrote the newest one. The guardrails above still hold. The human right-of-way rule still stands: a teammate who replies first stops the agent.

Teach it your business

Instructions tell an agent how to behave; knowledge tells it what is true. The workspace library lives at Agents → Knowledge. It lists every source, its status, and a short summary of what your agents know. You can add two kinds of sources, shared across your whole workspace:

• Documents — upload a PDF, Word file, spreadsheet, or plain text (price lists, policies, FAQs, service menus).
• Each file tops out at 15 MB.
• Your website — give it your site's address.
• CommSync reads up to 30 public pages of the same site.
• It reads only the pages that the site's robots.txt allows.
• The agent can then answer from the same information your customers see.
• A workspace holds up to 20 sources.

Once a source shows Ready, turn it on in the agent's Tools page. From then on, the agent looks up facts in its attached sources before it answers. When it cannot find an answer there, it escalates to a human rather than guess. Each agent only answers from the sources you attach to it, so a billing agent and a support agent can work from different material.

Connect the tools you already use

Beyond knowledge, agents can use external tools — services your business already runs, like a project tracker, calendar, or ticket system, connected over MCP. Connections belong to your workspace, not to one agent. Set a server up once in Agents → Tools, under Apps (also under Settings → MCP). Then turn it on for every agent that needs it. Ask can use workspace connections too.

• Start from Add an app. The list shows a tile for each app that runs its own server, such as Notion.
• Press Add on a tile to connect it, or open the tile to read about it first.
• Search finds any server that CommSync checked. Each app shows once.
• Other developers build their own versions of popular apps.
• Those versions show only on the page of the app, under Other versions.
• So the list never holds five copies of one app.
• When the app's maker runs no server, the app's page lists the versions by other developers, and says so.
• Browse the catalog at Agents → Catalog, by category.
• Add a server by address when your own team runs one.
• A custom server can carry an access token if it needs one.
• CommSync stores that token encrypted and never shows it again.
• Need two accounts on the same service? Connect the server twice.
• Turn on only the tools you want. A new connection turns on no tools.
• Open the connection, and CommSync lists every tool the server offers, each with a switch.
• Ask and your agents cannot see a tool that is off.
• A new tool on the server stays off until you turn it on.
• Turn a connection on for an agent. Use the Agents list of the connection, or the agent's Tools page.
• A workspace holds up to 25 connections, and one agent carries up to 10.
• One turn mounts at most 30 external tools.
• A receptionist and a dispatcher can use different tools from the same workspace.

Two things stay true no matter what you connect:

• The agent can never message anyone through an external tool.
• Tools let it look things up or act on other systems.
• Every reply to a contact still goes through its normal reply pipeline.
• That pipeline includes the approval queue in draft-first mode.
• CommSync records every external tool call on the turn's Activity trace.
• You can always see what the agent looked up, and when.

Call your own APIs

An agent can also call your own HTTP endpoints, with no MCP server needed. Describe an endpoint once in Agents → Tools, under Your APIs. Test it there, then turn it on for the agents that need it. See Agent webhook tools.

Every turn is auditable

Each time an agent acts, or deliberately declines to, CommSync records the turn on the agent's Activity tab. The record shows the message that triggered it, what the agent looked up, what it drafted or sent, and why it skipped a turn. If a reply ever surprises you, the answer to "why did it say that?" is one click away.

In the conversation itself, a reply that an agent sent shows Sent by and the agent's name under the bubble. So you can always tell an agent's reply from a reply that a teammate typed. Every teammate who can see the conversation sees the name, also a member who cannot open the agent settings. A reply that a person approved from the approval queue shows the agent too, because the agent wrote it.

Your daily brief keeps you posted too. On days your agents were active, the brief includes a short stat line. It shows how many conversations they handled and how many drafts sit in the approval queue. Quiet days add nothing.

After a successful send, the agent keeps a short private memory for that contact: durable preferences, commitments, and open items rather than a verbatim transcript. It uses that summary in later conversations with the same contact.

Test an agent

A test checks how an agent answers, and sends nothing. Open the agent's Tests page and select New test. Then pick one of three kinds:

Each result comes with a reason in one sentence.

Select Run all to run every test. To see how steady an agent is, open the menu next to Run all and run each test 3 or 5 times. The page shows how many runs passed. It also groups the reasons that tests failed, for example "Escalated instead of looking up invoices", so one fix can clear several failures.

To make a test from a real conversation, open the agent's Activity page, expand a turn, and select Make a test from this. The test starts with a copy of the conversation and the reply that the agent wrote then. You can use that reply as a good or a bad example.

What a test does:

• It sends nothing. It changes no conversation, no memory, and no contact.
• A tool that looks something up runs for real.
• An app tool does not run. It returns a stand-in result.
• One of your APIs with Retry if it fails turned off does not run. It returns a stand-in result.
• It uses AI credits, like a real reply. The page shows what a run cost.
• It runs on a paused agent too.

An agent can have 50 tests. One run holds up to 150 test runs (tests times repeats), and an agent runs one at a time. An AI app can do all of this over the MCP server.

Analysis

Analysis tells you if an agent does its job. When a conversation ends, AI reads it once. It answers your goals ("Did it book a visit?") and captures your fields ("Visit date", "Budget"). It also writes a title, a short summary, and how the contact felt. Open the agent's Analysis page to set it up.

Conversations

An agent conversation is one stretch of talk with a contact. It ends after a quiet time with no new message: 1 hour for a text and 1 day for an email, by default. You can set each time in Settings on the Analysis page, from 15 minutes to 7 days.

When the same agent answers a contact by text and by email, the two can be one conversation. That happens when a workspace shared contact holds both the number and the address. A teammate's private contacts never join conversations.

Each conversation also gets an outcome, which CommSync works out with no AI:

Goals

A goal is a question about the conversation. Each goal has these settings:

Each result is met, not met, unknown, or does not apply. Unknown means that the conversation does not show it either way. Each result comes with a reason in plain words and the quotes behind it. The success score of a conversation is 0 to 100, weighted over the goals that applied. An unknown result does not count.

An agent can have 30 goals. When you change a goal, results from before the change stay as they were and count apart from the new ones.

Fields

A field is a detail to capture. It has a type:

• Text, a list of text, an email address, or a phone number.
• A number, a whole number, or an amount of money.
• Yes or no, or one of a list.
• A date, or a date and time.

AI captures a value only when the conversation states it, and keeps the quote it came from. A value of the wrong type stays empty. A field never captures a password or a one-time code.

Mark a field sensitive to keep it out of webhooks, and away from an AI app that you limit to some lines. An agent can have 40 fields.

Try it, and templates

Try it runs a goal or a field on the agent's last 5 ended conversations before you save it, and shows each result. It saves nothing. It uses AI credits.

Templates add a starting set: General, Front desk, Scheduling, Billing questions, Support, and Sales leads. A template adds only the goals and fields that the agent does not have yet, and each one stays editable.

Alerts

An alert tells you when a conversation of an agent needs a look. Set the alerts on the agent's Analysis page. There are three kinds:

When an alert fires, every owner and admin gets a push notification. Turn on Also send an email to send them an email too. An alert never sends a message to a contact.

CommSync checks each conversation when its analysis ends. It checks again when a person corrects a result. One conversation fires one alert only one time, also after you analyze it again.

An agent can have 20 alerts. Turn an alert off to stop it and keep it. When you delete the goal or the field of an alert, the alert cannot fire.

Conversations to review. The Review page also lists the conversations to review, each with its reason:

• The agent did not meet a critical goal.
• A person gave a reply a thumbs down.
• An alert fired.

Open a conversation from the list to see it. Select Mark reviewed to remove it from the list.

The Overview

Analysis → Overview shows how your agents do, in numbers. Pick a range of 24 hours, 7 days, 30 days, 90 days, or 12 months. Filters narrows the numbers with the same filters as the Conversations view. Two dates in the filters make a custom range. A member sees only the conversations on lines they hold.

Select a number, a part of a goal bar, a value, or a row to open the conversations behind it.

A person's correction counts in place of the AI's answer. Agreement with the AI compares the two. It counts the goal results on the conversations that someone marked reviewed, and every corrected result. A goal counts only the results of its current version. Turn on Include earlier versions of a goal to count the older results too.

The values of a sensitive field never show on the Overview. It shows only how often the agent captured the field.

Weekly email. An owner or an admin can turn on the weekly email at the foot of the Overview. Each Monday at 9 AM in their own time zone, every owner and admin gets the numbers of the last week by email. The email holds numbers only, never message text. It is off by default.

Topics

Topics show what contacts talk to an agent about. Each night, AI puts the analyzed conversations of the last 30 days of an agent into 3 to 12 topics. For example: Visit times, Prices and fees, Changes to a visit.

• An agent gets topics when its analysis is on. It also needs 10 analyzed conversations in the last 30 days.
• AI reads only the title and the summary of each conversation, never the full messages.
• Each night's run replaces the topics of the agent.

The Topics table on the Overview shows each topic with its number of conversations, the moods, and the success score. Select a row to open the conversations of that topic.

An owner or an admin can build the topics of an agent now, and not wait for the night. Each run of the topics uses AI credits.

The Conversations view

Analysis → Conversations, in the Agents sidebar, lists every agent conversation of the workspace. A member sees only the conversations of agents whose lines they hold.

The filters:

Select rows to mark them reviewed, or to analyze them again.

Open a conversation to see its summary, each goal with its reason and its quotes, and each captured value. Select a quote to jump to its message. The transcript shows what the agent did between the messages. That includes its lookups and a handoff. It also includes a draft that the safety check held, or that a person approved or edited.

What you can add:

• Correct a goal result or a field value. The AI result stays visible next to it.
• Your correction counts in the success score, the filters, the export, and the webhook.
• Give each agent reply a thumbs up or down.
• Give the conversation a rating from 1 to 5, and a note for the team.
• Mark reviewed.

A conversation needs review when the agent missed a critical goal, or when someone gave a reply a thumbs down. It stays in Needs review until someone marks it reviewed.

Settings and cost

• A new agent starts with analysis on, at Standard depth, with the template of its kind. An agent from scratch starts with the General template. An agent made before Analysis keeps it off until you turn it on. Analysis reads only conversations that end after you turn it on.
• Depth. Standard costs about $0.10 to $0.30 for every 100 conversations.
• Thorough uses a stronger model. It costs about $0.40 to $1.00 for every 100 conversations.
• Share to analyze. Read all conversations, or a share of them, to lower the cost on a busy agent.
• Skip short conversations with fewer messages from the contact than you set.
• You can turn off Summary and title, and Mood and frustration.

Each analysis shows as "Agent analysis" in your AI usage. Analyze again on a conversation runs it once more, for example after you change a goal.

Security messages and one-time codes never reach the analysis, the same rule as every agent read. A webhook endpoint can receive each result as the agent_conversation.analyzed event. See Webhooks.

History

Every save of an agent is a version. Saves by one person less than 15 seconds apart make one version, so one Save is one version. A save of the setup, the rules, the reply style, the settings, the lines, the tools, the tests, or the goals makes one. Select History at the top of the agent to see them, the newest first. Each row shows the version number, who saved it, when, and what changed, for example "Changed the instructions and added a rule".

• Restore brings back an older version, after one confirmation.
• A restore makes a new version, such as "Restored version 3". You can undo a restore too.
• Restore brings back the setup, the rules, and the settings.
• Lines, tools, tests, and goals stay as they are.
• Each reply and each test run records the version it ran on.
• The conversation and the test results show Version and the number.
• An agent made before History starts its history at its next save.

Plan limits

Agents are a plan feature with one limit: how many agents your organization can have. The cap is org-wide, not per seat.

There is no monthly cap on how much your agents can do. Agent replies are usage-based. Each time an agent actually engages the AI to handle a message, that run draws from your workspace's pooled AI credits. That is the same allowance that powers labels, the Daily Brief, and Ask. The guardrail checks that happen before the model — skipped senders, quiet hours, opt-outs, human right-of-way — never reach the AI and cost nothing.

If the workspace's included credits and purchased balance both run out, agents pause along with the rest of CommSync AI. They resume when credits refresh or a top-up lands. The workspace monthly spend cap pauses agents the same way. When the workspace reaches the cap, agents stop for the rest of the month, together with every other AI feature. The agent's Activity tab shows those skipped turns with the reason Workspace monthly AI spend cap reached. So you can tell a cap from an empty balance. No message disappears: every one still lands in the inbox as usual for your team to handle.

A reply an agent sends counts as your reply on Home. When the agent answers an ask, the to-do for that ask closes as soon as CommSync sends the reply, as if you had replied.

Analysis → Overview, in the Agents sidebar, shows each agent's replies and cost next to its results. See The Overview.

Other limits:

How to set up an agent

What the draft holds

• Name, role, and instructions. The instructions say what the agent handles, what it must not do, and when it hands the conversation to a person. The draft uses only the facts you wrote. Where you gave no hours, prices, or addresses, it tells the agent to share your listed details. Add them to the instructions, or to the agent's knowledge.
• Tests. Each test is one message from a contact and what a good reply does. One test checks that the agent hands a hard case to a person. Run them on the agent's Tests page before you turn it on.
• Goals. One or two goals for the job this agent does, such as a booked visit. The agent also gets the starter goals and fields of its kind (Front desk, Scheduling, Billing questions, Support, Sales leads, or General). See Analysis.

To skip the draft, select Start from scratch. You then write the name, the job, and the instructions yourself. Write with AI drafts the instructions from the name, the job, and a short brief. When the instructions already exist, the same button reads Improve with AI. It uses a small amount of AI credit. An agent from scratch starts with the General goals.

An AI app makes the same draft with the MCP tool draft_agent. See MCP.

Reply style

The Setup page sets how long the agent's replies are on each channel, under Reply style.

A text is always plain text, so texts have a length only.

The agent's picture

Every agent gets a generated portrait. To use your own image, open the agent's Setup page and click the avatar next to its name. Upload a PNG, JPEG, WebP, or GIF image of up to 2 MB. To go back to the generated portrait, remove the picture.

The "Sent from CommSync" footer

Each agent has its own switch for the Sent from CommSync footer. Find it on the agent's Setup page, under Tone, language, and signature. The switch is on for a new agent. When it is on, every email the agent sends ends with the same small Sent from CommSync line, after the agent's signature. Turn the switch off to send without the branding.

Your own account has the same switch, and it adds the same line to your email. Find it under Settings → Account, in the section for email preferences. The footer applies to email only. An SMS never carries it.

An agent created before this switch existed keeps the footer off until you turn the switch on.