Manthan — Operating Principles
These principles are the platform. The design exists to honour them, not the other way around.
1. Exploration, not action
Manthan agents are always-on, autonomous, generative. They debate, prototype, critique, synthesise.
They do not ship to production. Action belongs to Conduir / Symposium / human operators. Manthan generates the options those systems then act on.
This is the cleanest line in the architecture. Anything that would let a Manthan agent reach into a production surface — file a real PR, run a real deploy, post to a real customer channel — violates this principle.
2. Noise is the substrate, not the enemy
We do not design rigid promotion gates / cull rules upfront. The process to extract signal from noise is itself an evolving research output of the platform — not a precondition.
"I am happy with noise, we will find the process to find information from noise. What is mess vs. organized disorder... let's start and then we will review and innovate and evolve."
The name is the principle. Samudra Manthan — gods and demons churn the ocean of milk together. Halahala (poison) surfaces first; then the 14 ratnas (gems); finally amrita (the nectar of immortality). Bad ideas come early. The rare breakthrough comes late. You keep churning. Adversarial and cooperative agents both contribute to the churn.
Pre-built gates encode current assumptions about what "good" looks like. Manthan's job is to surface where those assumptions are wrong. The cull function is itself an evolving artefact — discovered by running, not specified upfront.
When tempted to pre-specify "the right cull rate" or "the right promotion gate" — resist. Ship raw and iterate.
The 1000-ideas → 200-prototypes → 1-2-opportunities ratio is a retrospective observation, not a throughput target. Actual ratios will emerge.
3. Totally independent
Manthan is its own system. Own repo (manthan/), own infra (Postgres, NATS, services), own docker-compose stack. No runtime dependency on maestro-platform.
The maestro-platform swarm builds Manthan. That is the only direction the dependency runs. Manthan agents do not import from maestro-platform; Manthan services do not call maestro-platform endpoints.
4. Read-only outward
External systems read Manthan via a stable public API: /ideas/search, /prototypes/list, /opportunities/current. They do not write into it.
A future v2 may add a feedback channel where downstream outcomes ("this idea was used in mission M, success/cost/latency = ...") flow back as fitness signal. Until then, the boundary is one-way.
5. Honest expectations
- Autonomous swarms generate volume more reliably than quality. Selection function quality determines value.
- LLM agents recombine; they do not invent unexplored primitives.
- Always-on cost is real. Wake/sleep cycles, per-agent compute caps, dormancy-on-empty-inbox required from day 1.
- Long-lived agents drift. Identity reset rituals / memory pruning / generations are needed by the time agents have weeks of context.
- True novelty is not on the menu. Manthan generates options for human strategists.
The platform's value proposition is honest: reduce the cost of generating raw material; do not pretend to reduce the cost of judgment.
6. Bounded cost from day 0
Always-on token spend is the failure mode that quietly kills always-on agent platforms. Hard limits live in code, not in plans:
- Per-agent token cap (per hour).
- Per-day platform USD cap.
- Dormancy after empty-inbox timeout — agents stop thinking when nothing's there to think about.
- Wake on inbox arrival.
- Kill switch documented and rehearsed.
Defaults shipped in .env.example; tighten in dev, loosen with eyes open.
7. Observable, even messy
We are happy with noise if it is observable noise. The viewer must let a human watch the commune live and see why an agent said what it said. Structured JSON logs from day 0; metrics endpoint per service; lineage on every artefact.
If the swarm goes off the rails, we want to be able to see it happening, not learn about it from the bill.
Tensions these principles deliberately leave unresolved
- How do ideas become prototypes? No formal gate. Champions emerge or they don't. We watch.
- How are opportunities synthesised? Synthesiser agent role + human review. Process to be discovered.
- What's the "right" persona mix? Unknown. Start with three; observe; add or retire.
- How long should an agent live before it's reset? Unknown. Watch for drift.
- What's the right cost ceiling? Unknown. Start tight.
These are features of the design, not gaps. Resolving them too early is the failure mode.