A millisecond decision gate for agent fleets

The problem

Teams deploying agent fleets face a dilemma: let agents act freely and risk a bad payment, deletion or disclosure, or put a human in the loop and lose the speed they deployed agents for. Post-hoc observability dashboards only show the damage after it happened.

The idea

Why now

Sarvam Prism (4 Nov 2026) answers in 44 ms as an open-weight decision model, Anthropic's FlashProve (2 Nov) makes each decision's model identity provable, and Message Threads (2 Nov) keeps long agent sessions cheap. Millisecond in-path gating with receipts is newly buildable from open parts.

What it combines

Sarvam Prism's single-pass decisions (millisecond verdicts) + FlashProve's proof of model identity (receipts) + Message Threads' cheap long sessions (the gate sees full context without re-sending it). The mix matters because a gate that is slow kills agent speed, a gate without receipts cannot satisfy auditors, and a gate without context makes dumb calls.

MVP

Weekend scope: a proxy for one agent framework's tool calls, a decision model scoring each call against three written policies (no payments, no deletions, no external sends), and a receipt log. Deliberately skip: multi-framework support, policy learning, and human-escalation UI beyond a simple queue.

How it makes money

The moat

Observability incumbents cannot move in-path without rebuilding around decision models and tool interception — their architecture is log-and-display, and blocking actions would put them liable for the actions they block, a business-model shift they will not make in a quarter.

Distribution

B2B wedge: sell to platform-engineering teams running agent fleets, starting as a drop-in gate for one framework's tool calls, then expand across the fleet once the first team's incident count drops.

Why it wins

Observability tools (LangSmith, Langfuse and similar) log actions after the fact; this blocks them before they happen, at decision-model speed, with a receipt an auditor can verify.

Risks

Biggest risk: false blocks that stall legitimate agent work and get the gate disabled. The MVP de-risks it with shadow mode first — score everything, block nothing — and only enables blocking once the false-positive rate is measured below the team's threshold.

Build it with

• Sarvam Prism: open-weight decision model that answers in 44 milliseconds44 ms single-pass decision scoring for every proposed action
• Anthropic FlashProve: cryptographic proof of which AI model actually answeredCryptographic receipts for each gate decision
• Message Threads: Anthropic's public beta that stops agents resending the whole conversation every turnCheap full-context sessions so the gate judges with context

Repo to start from

gatekeeper-agents — a millisecond in-path decision gate for agent tool calls, with verifiable receipts.

Evidence

• Sarvam Prism launch
• Anthropic FlashProve
• Compliance cost and timeline
• Postman AI plugin leak discussion