Agent chat workspace

The workspace is a split screen: the agent on the left, your records live on the right. Ask it to look at a contact and the contact panel opens; ask for research and the Generate tab fills as the report is written. The agent reads the record you currently have open, so "draft something for this one" needs no further explanation, and its writes land in the same database you are browsing — not in a separate AI sandbox that syncs back later. Six tabs sit on the right: Contacts, Tickets, Campaigns, Knowledge, Generate and Deck. Under the message box are the Agent menu, a Knowledge switch, a Tone picker, attachments, a microphone and a language selector. Conversations persist as named sessions in tabs, so several threads run in parallel — prospecting in one, a single deal in another. Tool activity stays visible while the agent works, so a long request shows progress rather than a spinner.

Deep search

Deep search researches a company, a market, a person or a topic across many web sources and writes it up as a cited report in the Generate tab. It takes minutes, not seconds. A run fans out across several parallel research passes, and the number of web fetches per run is capped — that cap is what bounds the cost. Three behaviours are worth stating. A report always ships: if a run hits its budget you get a shorter report with an explicit Limitations section naming what could not be filled, never nothing at all. Citations are deduplicated and renumbered in order of first appearance, with the source list rebuilt from the citations that survived into the final text, so numbering and prose never drift apart. And fetched pages are handled as data, never as instructions — a page telling the agent to ignore its task cannot hijack the run. Public web only: no paywalled sources, no logged-in pages.

Scheduling jobs

A request can outlive the conversation that started it. Ask for something recurring — "every Monday, check these accounts for news and open a ticket if anything moved" — or something long — "research each of these 400 companies and log what you find" — and the agent stages a job for you to confirm. It then runs on its own, delivering results to the workspace, to Telegram, to Google Chat, or a combination. A job cannot start itself: a model tool call only stages it, and a human confirmation arms it, so a misunderstood instruction cannot quietly become a process that runs forever. For long lists, the rows never pass through the agent's context — the list is expanded and stored by the system and each item handed over on its own, which is why a job resumes cleanly after an interruption and why a 5,000-row list does not degrade as it goes. Recurring jobs run at a five-minute minimum; batch items drain one per fire, so a 500-item job is a multi-day process by design.

OUTREACH

Ask the agent to write to someone and it drafts the message grounded in what it knows about them — an opener, a connection note, a reply to a post. Two things do the heavy lifting. Every draft passes a voice check before it reaches you, flagging the tells that make a message read as machine-written: filler phrases, em-dashes, emoji, exclamation marks, the "not just X, but Y" construction. And there are deliberately no sample messages in the agent's instructions, because specimens get imitated and two different prospects end up receiving recognisably the same note. A draft is text, so it travels wherever you do — and today the one-click hand-off is built for LinkedIn, with connection notes held to the platform's 300-character limit before you ever see them fail. Other channels, email among them, are next. A contact marked do not contact is refused outright, and kōdlo never sends anything: the click is always yours.

Telegram & Google Chat Integration

Not a notification bridge — the actual agent. Ask it to research a company from your phone and it does, and the answer comes back in the chat. It runs the same pipeline as the web app, with the same access rules, the same allowance accounting and the same confirmation gate, so a change to your records proposed on your phone renders as tappable buttons and can be approved right there. Telegram connects through a one-time link in Settings → Connections; Google Chat connects the first time you message the bot. Generated documents arrive as files in Telegram and as a formatted card in Google Chat. Both connections are per person, not per workspace. Two honest limits: voice notes are not supported — the bot will ask you to type — and job cards do not render in either bot, so scheduling still gets confirmed in the web app.

Contacts

A contact record holds job title, location and country, email, phone, LinkedIn address, company name, website, industry, employee range, year founded, and free-text background covering education and career history. Each carries a pipeline status — one of eight, from contact through prospecting, engaged, nurturing, qualified and meeting scheduled to rejected and do not contact — and a dated timeline of everything that has happened, with workspace-defined event types you set up yourself. The record is the substrate everything else works on: Deep search writes its findings here, Discovery lands new prospects here with their evidence attached, and outreach drafting reads from here. One deliberate asymmetry is worth knowing: do not contact is a human decision. The agent cannot set that status, and cannot edit a contact already marked with it.

Generate

Reports, documents and drafts arrive in the Generate tab as files you can open, read, edit and keep, each with a stable link — and readable while they are still being written rather than appearing all at once at the end. The agent can read its own past output, which is what makes "update the brief we wrote last week" a real instruction rather than a request to start over. Documents can be edited in place, deleted and restored, or saved into the Knowledge Hub so they become citable reference material for later work. Deletion is soft; the underlying content is purged after 30 days, and asking the agent about a deleted document returns a clear answer rather than an error. These are Markdown-based text documents — a working format, not a page builder.