Autonomous agents call tools and models recursively, and a single loop can drain a budget or exfiltrate data before anyone notices.
Why this surfaced. Developers publicly report four-figure overnight bills from recursive agent loops, which is a visceral, budget-owner-level pain.
What the opportunity is, and why the buyer cares enough to pay.
A policy layer in front of agent tool calls that enforces spend ceilings, rate limits, allowed-tool lists, prompt-injection screening, and egress rules, with a kill switch and full audit trail.
Target customer
AI product teams, platform engineering groups, and enterprise security teams
India relevance
Indian services firms operating agents for global clients need demonstrable spend and data controls to win those contracts.
Why it matters
Agents blur the line between application bug and security incident, and existing APM tooling has no concept of an agent step.
Signal intensity over the trailing six months, and what drives it.
Runaway cost and prompt-injection threads dominate practitioner forums as teams move agents from demos into production.
Roles companies are already paying for to solve this internally.
| Role | Volume | Salary band | Postings |
|---|---|---|---|
| AI Security Engineer | Medium | ₹22L - ₹38L p.a. | 24 |
| Platform Reliability Engineer | High | ₹16L - ₹28L p.a. | 52 |
The expertise required, and how hard it is to hire in India.
Stops untrusted content from steering privileged tool calls.
Expresses spend and tool permissions as enforceable rules.
Rules that convert a nice-to-have into a mandatory purchase.
No regulatory driver identified. Demand here is market-led rather than compliance-led.
What recently became possible or cheap enough to make this viable.
| Technology | Adoption | Relevance |
|---|---|---|
| Model Context Protocol | Emerging | Standardises tool interfaces, which is exactly where a policy layer can sit. |
Who is already here, and what they are leaving on the table.
| Player | Category | Strength | Pricing |
|---|---|---|---|
| LLM gateways | Cost and routing proxies | Emerging | Usage-based |
| Cloud WAF vendors | Adjacent security tooling | Medium | Enterprise contracts |
The gap
Gateways handle cost. Security tools handle traffic. Nobody enforces both against the semantics of an agent step.
A concrete first version, and how it would make money.
Recommended MVP
A proxy SDK with a declarative policy file for per-run spend caps, tool allow-lists, and an injection screen, plus a dashboard replaying blocked steps.
Monetisation hypothesis
Per-protected-agent-run pricing, with an enterprise self-hosted licence.
What would most plausibly kill this. Read before committing capital.
Indian enterprises lack granular, self-hosted control over escalating LLM token costs, prompt injection risks, and shadow AI usage.
Buyer: CTOs and Engineering leads at mid-to-large Indian SaaS companies and BF…
IT Services & Enterprise Software
Indian IT services firms face severe margin compression when manually migrating enterprise legacy code (Java 8, PHP) to modern architectures (Go, Node.js, Python) with…
Buyer: Engineering VPs, Enterprise Architects, and Delivery Heads at Indian IT…
12 signals from github and rss clustered around implementation, india, parent, agent without matching any tracked opportunity.
Buyer: Not yet determined from the available signals