Custom AI Agents
Agents that finish the work.
Chatbots reply. Agents act. We build task-specific AI agents that research, draft, monitor, and execute — autonomously, on your real systems, with a complete audit trail you can review.
Audit log per run
Human-in-the-loop checkpoints
Hosted in Ireland / EU
Agent · Sales / Ops / Research
3 agents online
Tasks completed this week247
Draft follow-up for Bridgewater quote
Sales agent
done
4m ago
Summarise yesterday's 6 support tickets
Ops agent
done
12m ago
Personalised outreach: Morgan O'Hara
Sales agent
running
web.search"DUB tech VC 2026"12 results
linkedin.profile@morgan_oharaVP RevOps
crm.lookupmorgan@northway.iono match
Reconcile Stripe payouts → Xero
Ops agent
queued
Weekly competitor pricing scrape
Research agent
queued
Next run in 14m · Slack alert on failureAll systems normal →
First, the difference
Chatbots respond. Agents execute.
They sound similar but the architecture is different — and the outcomes are too. Knowing which you actually need saves you 80% of the project cost.
Chatbot
- Waits for a human message
- Replies in real-time conversation
- Lives inside one channel (web widget)
- Stateless or short-term memory
- Best for: lead capture, support deflection, FAQ
Agent
- Runs on a schedule or trigger — no human prompt needed
- Plans its own steps, calls tools, retries on failure
- Reaches across systems: CRM, email, sheets, APIs
- Long-term state + audit log of every action
- Best for: outbound research, ops automation, monitoring
What we build
Four shapes of agent.
Research agents
Scrape, summarise, enrich. Daily competitor sweeps, lead enrichment, market signal monitoring. Output drops into a CRM, sheet, or Slack channel.
e.g. Lead enrichment from a name + company
Outreach agents
Personalised cold email at human-grade quality. Researches the prospect, drafts in your tone, schedules sending, watches for replies, hands off when warm.
e.g. Weekly outbound to 50 fresh prospects
Ops agents
Triage, reconcile, route. Match invoices to POs, classify support tickets, route enquiries by intent, flag anomalies. Cleans the work before a human sees it.
e.g. Stripe ↔ Xero reconciliation
Monitoring agents
Watch a feed (logs, prices, mentions, sites) and act when criteria match. Less 'chat', more 'pager that thinks'. Run continuously, fire alerts or actions.
e.g. Brand mention monitor with reply drafts
Inside a single run
From trigger to logged outcome — fully visible.
Every agent run is a structured trace: goal in, plan generated, tools called, results captured. You can replay any past run, see exactly what happened, and adjust without rebuilding.
01 · Trigger
Trigger
Cron, webhook, manual button, or queued message. The same agent can be invoked many ways.
02 · Goal + context
Goal + context
Goal string + relevant data (lead row, ticket payload, scheduled job params) loaded into the agent's context window.
03 · Plan
Plan
Agent decomposes the goal into ordered sub-tasks. The plan itself is logged before execution starts.
04 · Tool calls
Tool calls
Each sub-task triggers tool calls — APIs, DB lookups, email sends, file writes. All results captured.
05 · Output + log
Output + log
Final output written to its destination (CRM row, email draft, Slack message). Full trace stored, replayable.
Built on
The agent stack we trust.
Claude (Anthropic)OpenAI GPT-4oLangChain / LangGraphn8nTemporalPostgresPineconeRedisBullMQHubSpot / Salesforce APIsSlack / Gmail APIsWebhooks
Where agents fail
Three ways an agent project goes wrong.
Vague goal
'Make us more productive' is not a goal — it's a hope. Agents need a measurable success criterion. The first session is always us narrowing down to something binary: did X happen or not?
No human checkpoint
Fully autonomous agents that send emails or move money in production are a disaster waiting. Every agent we ship has at least one approval gate for irreversible actions — even if it's just a one-line Slack confirmation.
No audit trail
An agent that runs daily but logs nothing is unmaintainable. When (not if) it does something wrong, you need to replay the exact run. Every Crescentek agent stores its full trace.
Frequently asked
Custom agent questions.
A daily prospect enrichment agent: takes a list of new leads from HubSpot, enriches each with LinkedIn + company data, drafts a personalised first-touch email, and queues them for human approval in Slack. ~2 weeks to build, saves a salesperson 4 hours a day.
Related AI services
Works alongside
What's the most boring 4 hours of your week?
That's usually where the first agent goes. Tell us the task — we'll scope what's automatable, what stays human, and how long until you're getting those 4 hours back.