01 / Autonomous agents
Agents that own an outcome, not a prompt
A workflow waits to be triggered. An agent decides. We build systems that run a business function on their own schedule: they find the work, research it, choose an action, take it, and follow up without anyone opening a dashboard.
Book a scope callWhat the system owns
- Discovery: the agent finds its own work from live sources, not from a static list
- Research: it enriches each target before it acts, and drops the ones that fail the bar
- Decision: an explicit policy layer with hard limits, not a model improvising
- Action: it writes, sends, files, posts or books, against the real API
- Memory: permanent per-target state, so it never repeats itself or contradicts itself
What gets shipped
- A model chain measured on your actual prompt, with fallbacks ranked by failure mode
- A durable state store so the agent survives restarts and knows what it already did
- Rate and spend ceilings enforced in code, not in a system prompt
- Run logs and a kill switch you control
Build shapes in this lane
03 patterns
Inbound qualifier
Reads every new lead as it lands, enriches it from the open web, scores it against your bar, drops the ones that fail, writes the survivors back into the CRM already researched, and starts the follow-up sequence itself.
Pipeline warden
Watches a set of long-running jobs, reads their logs, classifies each failure, retries what is retryable, escalates what is not, and files a written post-mortem against the run.
Research desk
Runs a standing brief on a topic or a market: pulls sources on a schedule, discards the ones that fail a freshness and credibility gate, and produces a dated, cited brief instead of a link dump.
Before you sign anything
- What stops it doing something stupid at 3am?
- Policy in code. Per-day ceilings, per-target one-shot rules, spend caps and a refusal list live in the runner, not in the prompt. A model can be talked into anything; a hard limit cannot.
- Which model does it use?
- Whichever one measures best on your actual prompt. Rankings do not transfer between use cases, so we benchmark the real prompt across a chain with repeats before anything ships, and we rank fallbacks by failure mode, because a loud error beats a silent empty answer.
- What happens when it breaks?
- It stops and tells you. Every gate fails closed, every run writes a log you can read, and there is a kill switch you own. Silent degradation is the failure mode we design against first.
Evidence
One of these is running right now.
@MiraAiHQ is a live autonomous operator built on exactly this architecture. The loop, the limits and the handoff are all documented.