Tamnoon introduced Tami, an agentic cloud‑remediation orchestrator that runs on Amazon Bedrock and moves the remediation workflow from manual, engineer‑heavy steps to an automated, safety‑checked pipeline. Practitioners care because the approach claims a >90 % reduction in critical exposure within 90 days while avoiding production incidents, which directly addresses the chronic overload of security findings.
Agentic pipeline on Amazon Bedrock
The system decomposes the remediation lifecycle into four specialized agents, each invoked as a serverless function and powered by Bedrock models through a single API endpoint. The agents operate independently, allowing each stage to scale and fail safely.
- Investigation agent consumes normalized findings, draws on a training corpus of real cloud environments, and produces a read‑only investigation plan. A LLM‑as‑a‑judge layer validates the reasoning before the plan is emitted.
- Safety‑scoring agent assembles an evidence chain per asset, merges automated verification results with analyst input, and classifies the remediation as
SAFE,RISKY, orAWAITING DATA. The model is deliberately biased toward conservatism, downgrading confidence when data is ambiguous. - Remediation agent activates only for findings the safety agent marked
SAFE. It orders steps, respects dependencies, and outputs a plan that follows a backup‑remediate‑verify pattern, including explicit rollback guidance. - Code‑generation agent translates the remediation plan into executable artifacts. An input guardrail screens prompts for relevance and protects against prompt injection. A separate reviewer agent must approve the generated code before it proceeds.
Safety scoring and guardrails
The safety‑scoring stage acts as the primary risk filter. By requiring a conservative classification before any write operation, the pipeline reduces the chance of production impact. The code‑generation guardrail adds a second layer: every artifact passes through a generate‑and‑review loop, ensuring that only vetted code reaches execution. These controls are explicit design choices rather than implicit trust in the LLM output.
Implementation and operational implications
Because each agent runs in an event‑driven, serverless environment, the architecture can handle bursty finding volumes without over‑provisioning. Normalization of inputs from CNAPPs, CSPM tools, and other scanners into a unified schema simplifies downstream processing but also introduces a dependency on consistent schema evolution.
Operationally, the pipeline introduces new observability requirements. Engineers must monitor agent health, model latency, and safety‑score distributions to detect drift or mis‑classification. The single‑API Bedrock integration removes the need for per‑model SDKs, but it also centralizes model access, making IAM policies around Bedrock calls a focal point for security reviews.
From a governance perspective, the explicit rollback steps and evidence chains provide audit trails that can satisfy compliance checks, yet they also require storage and retention policies for the generated artifacts and scoring metadata.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Adopting an agentic remediation pipeline on Bedrock suggests several concrete actions:
- Map each remediation sub‑task to a model that balances reasoning depth and cost; Bedrock’s single API makes this mapping straightforward.
- Implement a conservative safety‑scoring layer that defaults to
RISKYwhen data is incomplete, and treat the score as a gating condition for any write operation. - Design serverless functions with idempotent handlers and clear rollback paths to satisfy the backup‑remediate‑verify pattern.
- Instrument end‑to‑end tracing for each agent to surface latency spikes, model errors, or unexpected classifications.
- Review IAM permissions around Bedrock calls and artifact storage to keep the attack surface minimal.
By treating investigation, safety validation, remediation planning, and code generation as distinct, observable services, teams can scale remediation without sacrificing production stability.
