Five autonomous agents in production, each with a human where it matters

01

The situation

More than a chat box on a model

Businesses need software that watches their inboxes, documents and data, acts on its own, and knows when to stop and ask a person.

• Support spread across email, SMS, social and chat
• Specialist work (legal research, analysis, log triage, care notes) that has to be right
• No room for an agent that acts without a check

02

What we built

One agent per job, with the guardrails that job needs

Each agent is built around a specific job, with its own tools, data and guardrails, and ships as a desktop app or runs on a server, on Claude models.

• CX Agent: watches email, SMS, Facebook and Instagram, and website chat; answers automatically, holds social replies for approval and escalates what it cannot resolve
• Legal Agent: case-law research with good-law checks, drafting, billing with VAT and trust accounts, conflict and KYC checks, deadlines and intake
• Data Analytics Agent: pre-registers causal tests before looking at data, and reports only numbers backed by a logged query and an adversarial audit
• SOC Agent: a Level-1 analyst that parses security logs, maps events to MITRE ATT&CK and recommends triage actions
• Daisy, built for Blackroses Support C.I.C.: care workers dictate or type notes on WhatsApp, Telegram or desktop, and Daisy writes, files and emails the care record
• The same agent work powers our own products, Orchestra and OktoPAs, and the always-on business assistants we run for clients around the clock

03

How it runs

How they run day to day

• Approval queues hold anything public before it goes out
• Audit trails record what each agent did, including a hash-chained ledger for the legal agent
• Scheduled work runs on its own; the agent escalates when it is not confident

Built with