Security and deployment
What is a pluggable runtime for AI agents?
Alexandru Lupascu, CTO & Co-founder · Updated
A pluggable runtime for AI agents is a layer that attaches to the harness you already run, such as Claude Code, OpenAI Codex, Cursor or your own agent loop, and decides on every turn what the model sees and what stays out. It does not host your agent and does not change your model. That makes it a different thing from a hosted agent runtime: the AWS documentation describes Amazon Bedrock AgentCore Runtime as "a serverless hosting environment" where "the agent logic is code you write". A hosted runtime answers where the agent runs. A pluggable runtime answers what the agent sees. You can use either one alone, or both together.
What happens on each turn inside a pluggable runtime?
A pluggable runtime wraps the model call your harness already makes: it builds the context before the call and keeps what was learned after it. elephant.broker, which uses the term for its product, lists seven steps per turn:
- Pre-turn hook: intercepts the turn before the model sees it.
- Assemble: pulls task-relevant context across tools, teammates and sessions.
- Rank and prune: keeps current, verified context and drops stale or duplicate facts.
- Inject: places the working set where the model pays attention.
- Model turn: your model, your harness, unchanged.
- Post-turn hook: captures decisions, corrections and outcomes.
- Write back: makes the result available to the next turn.
"Pluggable" means the layer attaches through extension points instead of replacing anything. Harnesses already expose such points: Claude Code's UserPromptSubmit hook, for example, runs "when a prompt is submitted, before Claude processes it" and can add context. The trade-off is extra processing on every turn.
How does it differ from a hosted agent runtime, a memory library and a harness?
They answer different questions: who runs the loop, where code runs, where facts live, and what the model sees this turn.
| Category | What it does | What stays yours | You need it when |
|---|---|---|---|
| Harness (Claude Code, Codex, Cursor, your own loop) | Runs the agent loop: prompt, model call, tool calls | Rules files, permissions, model choice | Always: it is the agent |
| Hosted agent runtime (Amazon Bedrock AgentCore Runtime) | Hosts and scales your agent code, with isolated sessions and built-in authentication | Agent logic and loop | You deploy custom agents as a cloud service |
| Memory library or provider | Stores facts and returns them when your code asks | What to store, retrieve and place | One agent must recall facts across sessions |
| Pluggable runtime | Decides per turn what enters the context window, then writes outcomes back | Harness, model and hosting | Several agents or teams need the same current context, with a record of why |
In AWS terms, a harness is where "the orchestration loop itself is provided", and Runtime hosts it.
When do you need a pluggable runtime, and when is it more than you need?
You need one when the problem is what the model sees, not where the code runs. Typical signals:
- several agents or harnesses work on the same business, and each starts from zero;
- two sources disagree, and nobody can tell which one the agent used;
- a security review asks for a record of what each agent read and why;
- context keeps growing, and cost and errors grow with it.
It is more than you need for one person in one tool with short sessions: the harness's own rules files and memory cover that. If recall across sessions is all you are missing, compare options in how to choose memory for AI agents. If you need to deploy custom agent code as a managed service, that is a hosted agent runtime's job, and a pluggable runtime does not replace it. The wider discipline is context engineering.
How does elephant.broker fit, and where are its limits?
elephant.broker is a pluggable runtime for AI agents that sits between your harness and your business, keeps pointers to your systems rather than copies, and does not host agents or replace your model or harness. Connection is through an MCP server, native plugins or a REST API, and every turn recalls from, and writes back to, one graph and vector store, gateway-isolated per tenant. elephant.broker reports on its website 40-80% less context per task, with the condition that "only verified, relevant knowledge reaches the window". That is elephant.broker's own reported result, not an independent or head-to-head test. Security and deployment questions are covered in what security teams should check.
Sources
- elephant.broker: overview and reported results, how a turn flows, security and deployment
- AWS: Host agent or tools with Amazon Bedrock AgentCore Runtime, AgentCore harness vs. Runtime
- Anthropic: Claude Code hooks reference
Product details come from each vendor's public documentation as viewed on 2026-10-23 and may have changed. Product names are trademarks of their respective owners; elephant.broker is not affiliated with them.