poiesis · ai · agents
Agents,
tuned to your work.
A boutique platform of pre-tuned AI agents. Each one comes with chosen model, evaluated prompt, and configured tools — so you skip the work of getting it to actually work.
— four agents
A small, tuned roster.
Not a general-purpose assistant. Each agent has been picked, prompted, and evaluated against a fixed test set.
— measured, not claimed
Every agent beats baseline.
Each Poiesis agent is run against the same task suite as a vanilla GPT-4 + generic-prompt baseline. Quality scored 0–1 by an LLM-as-judge with a cross-vendor verification pass.
— architecture
How it works.
- 01
Curated layer
Each agent has a chosen model, an evaluated prompt, and a fixed tool set. Picked once, audited often. read more →
- 02
Runtime
Streaming, long-running tasks, sandbox isolation, cost monitoring — common plumbing, solved once. read more →
- 03
Channel adapter
One agent, three surfaces — Web, Telegram, and Lark — sharing identity and history. read more →
- 04
Eval harness
A fixed test set per agent, scored by LLM-as-judge with cross-vendor verification. read more →
— automation
Put agents on a schedule.
Any agent can run on a cadence and deliver the result where you already work — no prompting required after setup.
- 01
Set a cadence
Daily, weekdays, weekly, or every few hours.
- 02
Deliver anywhere
Web, or a paired Telegram or Lark bot — several at once.
- 03
Always recorded
Every run appears in your traces and counts toward the daily cap.
— pricing
Pay for what you run.
Start on pay-as-you-go, or include monthly credits with a plan. Every run draws from a visible balance — no surprise bills.
Pay as you go
- No monthly fee
- All four agents included
- Top-ups at $0.03 per credit
- Credits never expire