A production framework for AI agents.
The Agent Framework is an 11-part operating model for taking AI agents from demo to production. It groups the work into Build, Run, and Govern so teams can design the agent, operate it safely, and prove what happened when it acts.
From framework to delivery system.
The source framework names the right production concerns. This adaptation keeps the structure, then turns it into a path an engineering team can work through in order.
Inspired by Sandeep Mehta and the 11factor.ai framework.
Build
Design and Development. The structural choices that make an agent portable, grounded, useful, and maintainable before it ever reaches production traffic.
Model Serving Layer
Put every model provider behind a stable serving contract so the product can route, fail over, compare, and upgrade models without rewriting feature code.
Read factorContext Management
Design the full context pipeline: instructions, task state, retrieved evidence, user history, tool results, compression, ordering, and expiry.
Read factorMemory Management
Give the agent deliberate short-term and long-term memory with retention rules, retrieval policy, ownership, and stale-memory controls.
Read factorIntegrations and MCP
Expose external tools, APIs, files, and data sources through a consistent integration layer that is discoverable, permissioned, and observable.
Read factorRun
Production Operations. The operating controls that keep agentic systems responsive, affordable, recoverable, and reviewable while real users depend on them.
Orchestration
Coordinate models, tools, workflows, agents, queues, retries, checkpoints, and recovery paths for complex tasks.
Read factorHuman in the Loop
Design approval, review, escalation, and override paths for decisions where autonomy is too risky or too ambiguous.
Read factorRate Limits and Latency
Keep agents responsive under provider limits, slow tools, concurrent users, long context, retries, and multi-step workflows.
Read factorCost Control
Track and shape LLM spend by task, model, tenant, user, feature, token type, retry path, and business value.
Read factorGovern
Risk and Quality. The quality, safety, and audit practices that make an agent trustworthy enough to improve rather than merely impress.
Evaluation and Observability
Measure quality and expose system behaviour through evals, traces, labels, dashboards, alerts, and regression checks.
Read factorSafety and Guardrails
Constrain agent behaviour with input handling, output validation, permissions, sandboxing, policy checks, and circuit breakers.
Read factorReproducibility and Audit
Capture enough state to replay, explain, review, and defend important agent decisions after the fact.
Read factorUse the framework as a production checklist.
Work through the factors in order, then loop back whenever a new model, tool, workflow, or risk tier changes the system.
Read the first factor