TK

Building Delyt - autonomous customer operations for lean teams · one agent, full customer lifecycle · shipping & learning in public from London

[TK @theagentmaster]

The "AI customer service" category is solving the wrong problem.

Support resolution is one step. The real work is everything before and after: qualifying the lead, activating the new customer, catching the churn signal before the renewal call.

An agent that only handles tickets is a very expensive Zendesk.

The category that's missing: autonomous customer operations. One agent. Full lifecycle.

[TK @theagentmaster]

@steipete Codex is growing and doing smarter works. It surprises me as I compare its performance to just a few months back.

Curious are you now the mind behind codex?

[TK @theagentmaster]

Most “AI agents in production” aren’t in production. They’re in demo mode with paying customers. The gap: real production agents handle multi-turn conversations, recover from tool failures, escalate when uncertain, and stick to your SOPs across 10+ turns. Most “agents” shipping today break by turn 4. Test yours. The drop-off is brutal.

[TK @theagentmaster]

Walking around SF as a founder right now is surreal. Two years ago the flex was headcount. Today it’s how much one person can ship with agents doing the work. The whole definition of “building a company” is being rewritten in real time.

[TK @theagentmaster]

@paulg 100%…Articulated in the best way. Hard not to ignore

[TK @theagentmaster]

@KaiXCreator Learn how to instruct it better. Writing code is no more the moat.

[TK @theagentmaster]

Our Sales Agent on Delyt isn’t a chatbot.

It prospects. Qualifies. Runs the conversation. Books the meeting. Closes when it can.
WhatsApp, email, live chat, Instagram, SMS - same agent, same playbook, every channel your buyer is on.
Quietly, the SDR role is changing.

DM to see a demo.

[TK @theagentmaster]

Agentic Publishing on @delytAi is now live 🚀 Our early customers absolutely love it.

Connect Delyt to Claude >> Ask Claude to research your competitors or what’s trending >>draft posts in your voice, and schedule on Delyt across LinkedIn, X, Instagram — all from Claude/Chatgpt.

Then when customers reply, Delyt’s agents handle the conversations for sales or support or ops etc to manage your social reputation and convert leads and help your brand grow in social.

Research → publish → engage → convert. One loop.

First platform to do this end-to-end.

DM and we will show you how. 🙌🏻

[TK @theagentmaster]

Most CX platforms are dashboards built for humans to click.
@delytAi is built for agents to operate.
Connect us to Claude. Ask Claude to handle today’s customer ops - qualify the new leads, follow up on stalled deals, resolve the 12 cases waiting on a refund decision or build a complex config or even spin few agents and assign them tasks. It can speak with delyt like a colleague and get things done that are unimaginable on legacy platforms.

No dashboard. No clicks. Just outcomes. Every action a tool. Every workflow callable. Every step auditable.
The interface to your customer operations is changing. We built for what’s next.

DM to see it in action.

[TK @theagentmaster]

Most 'AI for customer ops' tools are just ticket routers with a language model in front.

The actual work: qualifying the lead, activating the new customer, catching the churn signal, running the renewal.

That's four workflows. Most tools handle one.

[TK @theagentmaster]

@gopikl The silent reroute is the real signal. Three weeks of great vibes followed by zero follow-up questions usually means the ops team stopped trusting the output and started doing manual workarounds instead of pushing back. That's harder to detect than an outright rejection.

[TK @theagentmaster]

@Surreal_Intel @ithilgore @sama @OpenAI The 'put the tools down' framing is sharper than 'safety.' Safety implies restriction. Knowing when to stop is a competence. The agent that keeps going when it should pause isn't disobedient, it just doesn't have enough signal to know the difference. That's a product problem.

[TK @theagentmaster]

@hansel_hansl @ithilgore @OpenAI The data-class axis is the right design question. Most teams collapse it to sensitivity level, which doesn't capture the real risk. PII read vs PII write are different actions. The ceiling-plus-class combo you described handles the edge cases that a flat sensitivity tier can't.

[TK @theagentmaster]

@zoepark_builds 8 days from vibe code to prod is the real benchmark shift. A year ago that was a 6-week sprint with 3 PRs and a QA cycle. The tooling did catch up, but so did the tolerance for 'good enough for now.' Both matter.

[TK @theagentmaster]

@zoepark_builds The 3-day train time on a workflow that replaced an $8k/month hire is the stat that matters. Most teams spend 6 weeks on discovery and never ship. The velocity advantage isn't just the agent, it's not having 4 approval layers before you can move.

[TK @theagentmaster]

@chloetechai The 70% stat raises a bigger question: what are those agents actually doing. Cloud ops is one thing. The harder design question is what happens when a workflow hits an unexpected state and no human is watching. That's where most enterprise deployments stall.

[TK @theagentmaster]

@xiao18kuma @daniavitz Data quality as the real constraint is underrated. The routing gets the credit when it works. When it doesn't, it's usually the trust layer. Agents can route perfectly and still fail if the enrichment data they're acting on was stale three days ago.

[TK @theagentmaster]

@daniavitz CAC is a workflow problem and so is conversion. The 25% improvement isn't from better targeting, it's from agents that handle the routing, qualification, and follow-up without human delay. The lead that would've gone cold in 48h gets worked in 48 minutes.

[TK @theagentmaster]

Most companies think they have an AI problem.

They have a workflow problem.

The model is capable. The loop around it isn't.

What triggers it. What it does with the result. Who checks it. What happens next.

That's where the real work is.

[TK @theagentmaster]

@adam75563 @GroverLovesh @alexmorris10x Outcome validity is a different layer from infrastructure health. Uptime tells you the agent ran. Latency tells you how fast. Neither tells you if the customer got what they actually needed. Most observability stacks stop right before the metric that matters.

[TK @theagentmaster]

@BetterSayAJ State management is exactly the right frame. Retrieval is stateless: you fetch what exists. State management is stateful: you track what changed, what was agreed, what's active. Most agent memory architectures are just a retrieval layer pretending to be the second thing.