Qué construimos después.
Proponé ideas, votá las que necesitás y seguí su estado. Lo más votado sube en la lista.
Agents: Crawl a whole site or sitemap with include and exclude patterns
Chatbase crawls a website or sitemap and auto-resyncs weekly (https://www.chatbase.co/docs/user-guides/chatbot/data-sources.md); Voiceflow manages sitemaps with include/exclude patterns and a refresh schedule (https://www.voiceflow.com/docs/documentation/build/importing-data-sources.md). We allow 20 single URLs (AGENT_MAX_URLS) re-synced daily. A team moving from an old help center on another domain has hundreds of pages. First scope: add a site or sitemap URL, discover pages up to a cap (e.g. 500) with include/exclude path patterns, fetch through safeFetch with a crawl budget per cron tick, show the page list with status, and re-sync weekly.
Agents: Add a bulk re-run of saved tests across all agents from one screen
Fin lets teams test changes against many past conversations before turning them on (https://www.eesel.ai/blog/intercom-fin-ai-explained). Saved tests exist per agent (components/agents/AgentTests.tsx) but each agent must be opened one by one. First scope: a Tests overview on the agents home that runs all tests for the selected agents and shows pass or fail counts with a link to each failing case.
Agents: Show a preview of which knowledge source answered each reply
Fin shows the content used to produce each answer in the conversation view so admins can fix the source (https://www.intercom.com/help/en/articles/8205718-fin-resolutions). We cite sources to the visitor (lib/agents-semantic.ts) but the conversation view in components/agents/AgentConversations.tsx does not list which chunks were retrieved. First scope: store the retrieved source ids per assistant message and render them under the reply with a link to open the Docs page or file.
Agents: Add custom guidance rules that apply by audience or topic
Intercom Fin has Guidance, short rules that apply only when a condition matches, such as the customer's plan or topic (https://www.intercom.com/help/en/articles/10444459-fin-guidance). Here the only free-form text is one Instructions field plus procedures (lib/agents-procedures.ts), so every rule is always in the prompt. A small team with a paid plan and a free plan needs different answers about refunds. First scope: a list of rules with a condition text and an instruction, matched per message by keywords, and a count of how often each fired in the analytics tab.
Agents: Export and import an agent as a JSON file
Chatbase and Voiceflow let teams move agent configuration between workspaces (https://www.voiceflow.com/docs/documentation/deploy/environments/overview.md). Duplicate now copies an agent inside a workspace (lib/agents.ts duplicateAgent), but there is no file to back up before a big edit or to share a template. First scope: Export button in agent Settings that downloads instructions, knowledge selection, tools, handoff, guardrails and saved tests as JSON without secrets, and an Import button on the agents list that creates a new agent and lists what must be reconnected.
- 0
Agents: Index the 7-day stats queries and stop loading them in the Flow editor
listAgents (lib/agents.ts:157-176) aggregates agent_conversations with WHERE user_id AND created_at and agent_feedback with WHERE user_id, but both tables only have agent_id indexes. EXPLAIN QUERY PLAN shows a SCAN of each table, so the cost grows with every conversation. The Flow editor page (app/[locale]/app/flow/[id]/page.tsx:27-30) loads the full listAgents, stats included, just for a picker. First scope: a migration adding agent_conversations(user_id, created_at) and agent_feedback(user_id, agent_id), plus a light listAgentOptions for the pickers. Priority: medium. Effort: S.
Agents: Fall back gracefully on the public page and API when the model fails
On the Chat channel a model failure falls back to the keyword bot (lib/agents-chat.ts), but on /a/<handle> the visitor sees the owner-facing error asking to load credits in the Vercel AI Gateway or choose a model in Instructions > Model (reproduced in the benchmark with ai_no_credits). Fin and Chatbase never show configuration errors to customers. First scope: map provider errors to a generic visitor message on the public page and API, offer handoff with an email field, record the error so the existing alerts notify the owner, and optionally retry with a fallback model.
Agents: Export, import and duplicate an agent with its configuration
Chatbase lets you duplicate agents and Voiceflow exports environments (https://www.chatbase.co/docs/user-guides/chatbot/settings.md, https://www.voiceflow.com/docs/documentation/deploy/environments/overview.md). There is no way to copy an agent here: to try new instructions safely you rebuild it by hand, and there is no backup before a big edit. First scope: Duplicate in the agent menu, and export/import as JSON with instructions, knowledge selection, tools config, handoff, guardrails and tests (secrets excluded and listed as missing on import).
- 0
Agents: Add step-by-step procedures the agent follows for a specific task
Fin Procedures combine natural-language steps with deterministic if/else, code and approval steps (https://fin.ai/help/en/articles/13976187-fin-procedures-explained); Chatbase has Procedures (https://chatbase.co/docs/user-guides/chatbot/procedures/procedures-overview.md) and Voiceflow mixes playbooks with canvas workflows. Our agent has one instruction block, so a refund or reschedule flow depends on the model following prose. First scope: named procedures with a trigger description and ordered steps (ask for a field, call a tool, check a condition, hand off); the runtime injects the active procedure and tracks the current step per conversation, and Analytics shows completion per procedure.
- 0
Agents: Connect MCP servers as tools
Fin added MCP connectors including a Zapier MCP that reaches thousands of apps (https://fin.ai/updates) and Voiceflow has an MCP tool (https://www.voiceflow.com/docs/documentation/build/tools/overview.md). Our tools are the suite tools plus http_request with secret headers (lib/agent-tools.ts), so calling a third-party system means hand-writing HTTP calls. First scope: add an MCP server URL with an optional secret header (reuse agent secrets), list its tools, enable specific tools per agent, run them through the existing logging, dry-run simulation and approval flow, and block private hosts with safeFetch.
- 0
Agents: Show buttons and short forms in the chat to capture leads
Chatbase has Collect leads, Collect data, custom buttons and suggested messages as actions (https://www.chatbase.co/docs/llms.txt). Our agents ask for email and company in free text and then call crm_upsert_contact; visitors skip or mistype it and the SDR template loses leads. First scope: a collect_info tool the agent can call with fields (email, name, phone, company, a select) that renders a small form in the Chat widget and the public page, validates on the client, returns the values to the agent and saves them to the CRM contact, plus quick-reply buttons the agent can offer.
- 0
Agents: Score every conversation for sentiment and quality with AI
Fin rates every conversation with a CX Score without surveys and detects sentiment and urgency (https://community.intercom.com/product-updates/a-new-era-of-insights-has-arrived-10513); Chatbase shows sentiment and CSAT (https://www.chatbase.co/docs/user-guides/chatbot/analytics.md). Our Analytics tab (lib/agents-analytics.ts) has resolution, handoffs, tool errors, keyword topics and thumbs, which few visitors click. First scope: when a conversation goes idle, a cheap model call scores sentiment, whether the question was answered and a 1 to 5 quality score with a one-line reason; show averages and a worst-conversations list in Analytics and feed the existing alerts.
- 0
Agents: Run multi-turn simulations with an AI customer and replay real conversations
Fin Simulations play a full conversation with an AI customer, mock connector outputs and judge the result; Batch Test replays real conversations (https://fin.ai/updates/simulations-for-procedures, https://fin.ai/updates). Voiceflow tests support AI-simulated users with success criteria and assertions on tool calls (https://www.voiceflow.com/docs/documentation/measure/tests.md). Our tests (lib/agents-tests.ts) are one question plus an expected behavior, which does not cover booking or qualification flows. First scope: a test can be a persona plus a goal; an LLM plays the customer for up to N turns in dry run and the judge checks the goal and which tools were called; plus 'Replay last 20 conversations' that reruns real visitor messages against the current instructions and highlights answers that changed.
- 0
Agents: Add WhatsApp as a channel for the agent
Chatbase deploys agents to WhatsApp (with templates and Coexistence), Messenger, Instagram and Slack (https://www.chatbase.co/docs/llms.txt) and Fin answers on WhatsApp, SMS and Slack (https://fin.ai/updates). Agents only run on a Chat widget, /a/<handle>, the API and Flow (AGENT_USAGE_CHANNELS). For small teams in LatAm WhatsApp is where customers write. First scope: connect a WhatsApp Business Cloud number to an agent, map each phone number to an agent conversation, run runAgent on inbound messages, reply within the 24 h window, and hand off to the Chat inbox (shared with the Chat WhatsApp channel idea) with the same handoff rules.
- 0
Agents: Upload PDF, DOCX, TXT and CSV files as knowledge
Chatbase trains on PDF, TXT, DOC and DOCX up to 20 MB each (https://www.chatbase.co/docs/user-guides/chatbot/data-sources.md) and Voiceflow imports PDF, DOCX and CSV/XLSX tables where each row becomes a chunk (https://www.voiceflow.com/docs/documentation/build/importing-data-sources.md). Our Knowledge tab only takes Docs pages, up to 20 URLs and Chat saved replies (knowledgeConfigSchema in lib/agents-config.ts); there is no file input (verified in the browser). Price lists and manuals usually live in PDFs. First scope: upload files to an agent (reuse the Docs upload storage), extract text server side, chunk and index them next to URLs for search_knowledge, show size, status and re-upload, with a per-workspace size cap.
- 0
Agents: Show API and Flow usage in the agent list channels and 7-day counts
listAgents in lib/agents.ts only pushes 'chat' and 'public' into channels even though AgentListItem allows 'api', so agents used through /api/v1/agents or the Flow agent node (lib/flow-engine.ts agentNode, which also runs as channel 'api') look unused in the list. Fin's reporting breaks volume down by channel (https://fin.ai/pricing). Owners cleaning up agents could delete one that a Flow depends on. First scope: add 'api' when the agent has any api-channel conversation in the last 30 days and list referencing flows on the agent page, warning before delete. Priority: low. Effort: S. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Support secret headers on http_request so agents can call authenticated APIs
http_request only sends a fixed user-agent and content-type to allowlisted domains (lib/agent-tools.ts) and toolsConfigSchema has no place for credentials (lib/agents-config.ts), so it can only reach public endpoints. Fin connects to internal systems via Data Connectors or MCP with control over what data is exposed (https://production.fin.ai/procedures), and Zapier Agents act across thousands of authenticated apps. First scope: per-domain encrypted headers (for example Authorization) stored server-side and never shown to the model, optional named endpoints with a description and JSON schema so the model picks an endpoint rather than composing raw URLs. Priority: low. Effort: S. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Answer email threads from CRM inboxes as a channel, draft-first
Fin deploys to email, WhatsApp, SMS, Slack and voice (https://fin.ai/pricing), and Lindy drafts and triages email (https://docs.lindy.ai/features/inbox-management/email-drafting.md). Our channels are the Chat widget, the public page, the API and the Flow node (AgentChannel in lib/agents-config.ts, lib/agents-chat.ts); there is no email channel even though the suite already handles email in lib/crm-email.ts. Email is where most small-team support and sales actually happens. First scope: assign an agent to a CRM email inbox in 'draft' mode, where it writes a reply draft on each inbound thread for a teammate to send, with the same tools and logging as other channels. Priority: medium. Effort: L. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Alert the owner when thumbs-down, tool errors or step limits spike
Fin Monitors evaluate live conversations against standards you define and alert when a threshold is crossed (https://fin.ai/testing, https://fin.ai/pricing). We record ratings in agent_feedback, tool errors in agent_tool_calls and the step_limit status in runAgent (lib/agents.ts), but only handoffs produce a notification (lib/agent-tools.ts). A small team will not watch a dashboard daily; a push when something breaks (for example a Cal event type deactivated making every booking fail) is what keeps an agent safe to leave running. First scope: hourly cron check per agent for tool error rate, thumbs-down count and step_limit count over the last 24h vs threshold, sending one in-app and email notification with a link to the failing conversations. Priority: medium. Effort: S. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Add semantic retrieval and show source links in answers
Our search is in-memory BM25 over keyword tokens with a small ES/EN/PT stopword list (lib/knowledge-search.ts), so a question phrased differently from the doc, or in another language than the doc, often returns nothing and triggers a handoff. Fin answers in 45+ languages from the same content (https://fin.ai/capabilities). search_knowledge already returns a source for URLs but the visitor never sees which page an answer came from, which hurts trust and makes wrong answers hard to debug. First scope: hybrid ranking (BM25 plus embeddings via the AI Gateway, cached per chunk with a content hash), and an instruction plus UI to append one source link (Docs public URL or web URL) to answers that used knowledge. Priority: medium. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Give the agent the identified visitor's CRM, Cal and Track context
Fin personalizes answers using customer data and connectors (Stripe, Shopify, Salesforce) and can route by customer attributes (https://fin.ai/capabilities, https://fin.ai/train). Our prompt (buildAgentInstructions in lib/agents.ts) contains no visitor identity even when the Chat widget already knows the visitor's email, and the agent can only learn a contact's first name and lifecycle stage via crm_find_contact (lib/agent-tools.ts). Owning CRM, Cal and Track lets us do this without integrations: upcoming bookings, open issues the visitor reported, deal stage. First scope: when the Chat visitor has a verified email, inject a short read-only context block (name, company, lifecycle stage, next booking, open Track issues by that email) and add tools to list or cancel that visitor's own bookings. Priority: medium. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Turn unanswered questions into suggested Docs pages the team can approve
Fin reviews real conversations to recommend content updates and surface missing topics, approvable in one click from its Optimize dashboard (https://fin.ai/train, https://fin.ai/capabilities). We already log every search_knowledge call with its query and empty results in agent_tool_calls and have Docs in the same suite, but nothing connects them (lib/agent-tools.ts, lib/agents.ts). This is a suite advantage: the fix lands directly in the knowledge source the agent reads live. First scope: weekly job that clusters empty-result queries plus handoff reasons, drafts an answer from the teammate replies in handed-off Chat conversations, and creates an unpublished Docs page linked to the agent for review. Priority: medium. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Add an analytics tab with resolution, handoff, tool error and topic metrics
Fin reports resolution rate, involvement rate and CX Score per conversation, groups conversations with Topics Explorer and offers Monitors and scorecards (https://fin.ai/pricing, https://fin.ai/capabilities). We only compute conversations in the last 7 days and all-time thumbs totals in listAgents (lib/agents.ts); the editor has no analytics tab (TABS in components/agents/AgentEditor.tsx). A team cannot decide whether to widen the agent's scope without knowing how often it resolves vs hands off and which questions it fails. First scope: a 'Performance' tab with 30-day series of conversations, handoff rate, step_limit rate, tool error rate by tool, thumbs ratio, and top unanswered queries (search_knowledge calls that returned 'no relevant knowledge found'), all from existing tables. Priority: high. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Add draft vs published versions with history and one-click rollback
updateAgent in lib/agents.ts writes instructions, tools and guardrails straight to the live row, so every save in the editor goes to production channels instantly and there is no record of what changed. Fin Releases keep changes in a space separate from the customer-facing version until you publish (https://fin.ai/testing), and AgentKit offers full versioning (https://openai.com/index/introducing-agentkit/). Without versions you cannot tie a drop in thumbs-up to an edit or undo a bad prompt quickly. First scope: an agent_versions snapshot on publish, playground and tests run on the draft, a Publish button with a short diff, and a version list with restore; store version id on agent_conversations for later analytics. Priority: high. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Let owners require teammate approval before specific write actions run
Zapier Agents pauses at approval steps and waits for you to approve or stop, and Human in the Loop lets reviewers approve, decline or edit data before the action proceeds (https://help.zapier.com/hc/en-us/articles/41776074420493); AgentKit has a User approval node (https://openai.com/index/introducing-agentkit/). Our tools are either fully on or off (toolsConfigSchema in lib/agents-config.ts) and execute immediately. Small teams want the agent to draft a Track issue, a CRM update or a Flow run and let a person confirm, especially at the start. First scope: per-tool 'needs approval' flag, a pending_actions table, an in-app notification with approve/edit/reject, and the agent tells the visitor the request is being reviewed; executed result is posted back to the conversation. Priority: high. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Make the playground a dry run that simulates write tools instead of executing them
The playground channel runs the same tools as production (lib/agent-tools.ts only relaxes the email check for playground), so testing an SDR agent books real meetings and sends confirmation emails, creates CRM contacts, opens Track issues and enqueues Flow runs. Fin previews and Releases let you try changes in a space separate from what customers see (https://fin.ai/testing), and AgentKit has preview runs (https://openai.com/index/introducing-agentkit/). This makes people afraid to test, which is the opposite of what builds trust. First scope: a 'simulate actions' toggle on by default in the playground that returns realistic fake outputs for crm_upsert_contact, crm_log_note, cal_book_meeting, track_create_issue, run_flow and POST http_request, marked as simulated in the tool call log. Priority: high. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Add saved test cases with an LLM judge and rerun them before publishing
Fin offers Simulations where one LLM plays the customer and another judges handoff timing, content used and standards, with a saved regression library rerun after every change (https://fin.ai/testing); Lindy has offline evals with LLM-as-judge scorers over reference tasks (https://docs.lindy.ai/fundamentals/lindy-101/evals); OpenAI AgentKit ships datasets and trace grading (https://openai.com/index/introducing-agentkit/). We only have a manual playground (components/agents/AgentChat.tsx) and per-message thumbs; there is no test table or runner anywhere under lib/agents*. A small team will not trust an agent after editing instructions without a one-click 'did I break anything' check. First scope: an agent_tests table (user message or short script, expected behavior in plain language, expected tool or handoff), a 'Tests' tab that runs all cases through runAgent in dry-run mode, judges with a cheap model and shows pass/fail with the transcript, and 'save as test' from any conversation. Priority: high. Effort: L. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Fetch knowledge URLs on save and re-sync them daily from the cron tick
Saving a new URL only inserts a 'pending' row (syncKnowledgeUrls in lib/agents.ts) and searchAgentKnowledge only reads status='ok', so a newly added page is invisible to the agent until someone clicks refresh in the Knowledge tab. There is no scheduled re-crawl (vercel.json only runs /api/cron/tick and robots, neither touches agents), so prices or policies on those pages silently go stale. Lindy auto-refreshes its knowledge base every 24 hours (https://community.lindy.ai/x/support/htoqrkuf750a/understanding-resync-knowledge-base-action-with-dy), Fin syncs website and Zendesk content (https://fin.ai/train), and Zapier schedules daily/weekly syncs (https://help.zapier.com/hc/en-us/articles/29457465505805). First scope: kick refreshKnowledgeUrls for new URLs after PATCH, re-fetch URLs older than 24h from the hourly cron in small batches, and show 'last synced' plus an error badge in the Knowledge tab. Priority: high. Effort: S. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: Stop replying after handoff on public/API and let teammates answer there
On the Chat channel the agent goes silent after handoff (lib/agents-chat.ts checks wasHandedOff), but sendAgentMessage in lib/agents.ts never checks agent_conversations.handed_off, so on /a/<handle> and the API the bot keeps answering after it promised a teammate. Worse, components/agents/AgentConversations.tsx is read-only: the notification links to a transcript where nobody can reply, so the handoff is a dead end. Fin keeps the conversation owned and routes it to the right team with a reply path (https://fin.ai/capabilities). First scope: after handoff, store visitor messages without running the model, add a teammate reply box in the Conversations tab that streams into the public page via polling, and ask for an email at handoff time so the team can follow up offline. Priority: high. Effort: M. Compared with Intercom Fin, Zapier Agents, Lindy (plus OpenAI AgentKit as reference for evals/versioning).
- 0
Agents: AI agents that answer in Chat and take actions
Agents with instructions, knowledge (Docs, URLs) and tools across the suite (CRM, Cal, Track), with human handoff.