DeepSeek this week published an open-source agent harness layer: a modular system that organizes tools, memory, execution, sandboxing, storage and even temporary plugins. The project is released under an MIT license, and its core idea is summed up as “Everything is a Plugin.”
The DeepSeek Harness makes components around an agent — provider, storage, execution, and tools — pluggable and interchangeable. Notably, an agent can create a temporary plugin when it lacks a capability, use it, and then remove it. That pattern starts to form an initial self-extending loop: the architecture hints that future competitive advantage may lie in modular, open orchestration above base models.
Researchers discovered secrets in agents’ encrypted reasoning
Alongside DeepSeek’s release, researchers reported that agents exchange encrypted reasoning blocks between model calls and that parts of those traces can be recovered without breaking the underlying encryption. In public agent logs they found passwords, API keys, access tokens and personal information hidden inside these reasoning blocks, even when the visible conversation did not contain that data.
The researchers also found that these blocks can carry instructions that influence later steps, so hidden reasoning may not only record past activity but actively change an agent’s subsequent behavior.
Scale and a new observability layer
Scale adds another dimension. Simile AI — which grew out of Stanford’s 25-agent Smallville experiment — is now valued at $2 billion and aims to ultimately simulate eight billion people. The original Smallville experiment had 25 characters living, remembering, planning and forming relationships inside a simulated town. Simile’s team moved from that research to product by modelling agents from two-hour interviews with 1,052 real people, then expanding the dataset to 2.9 million responses from 400,491 participants across 210 social-science experiments.
As agents call tools, retrieve data, update memory and hand off tasks, simply monitoring model inputs and outputs is insufficient. When failures occur, engineers need to know whether the problem was in the model, the retrieval layer, a tool call, memory handling or the infrastructure. For that reason, agent observability is becoming its own layer of the stack.
Open-source monitoring and observability tools
There are multiple open-source tools aimed at revealing what agents actually do and helping debug them. Examples mentioned include Langfuse and AgentOps, alongside general observability tools such as OpenTelemetry, Prometheus and Grafana. These tools help trace which components were invoked, when, and where errors originated.
A quick refresher on LoRA
LoRA (Low-Rank Adaptation) offers a lightweight route to adapt very large language models: instead of updating all model parameters, the base model is frozen and only small, low-rank matrices are trained to capture task-specific changes. That approach reduces cost and memory footprint, makes switching between tasks easier and allows the trained matrices to be merged back into the model without adding inference latency.
Conclusion
The combination of DeepSeek’s open harness and the discovery of sensitive data inside encrypted agent reasoning highlights a tension: the infrastructure around agents is becoming more powerful and more modular, but also harder to inspect and secure. As systems grow in scale and autonomy — as illustrated by Simile AI’s ambitions — observability, traceability and security are rising to the level of core engineering concerns.
