September 2026 Summaries

No summary generated yet.

DeepSeek Harness (dsh) is an MIT-licensed, developer-preview agent runtime built on the Cordis plugin framework, where model adapters, tools, session storage, shells, subprocess execution, subagents, and even the agent loop can be replaced through configuration. The review finds that it can connect to DeepInfra and other OpenAI-compatible endpoints through custom provider settings, though reasoning models may require YAML compatibility overrides for request fields and roles, and credentials must be available in the environment that launches the process. Its browser UI is the default interface, but headless CI-oriented operation and community terminal UI profiles are available; it also supports session trajectories, sandbox modes, MCP, existing Claude Code and Codex hooks, and delegation to Claude Code, Codex, or Agent Client Protocol agents. The article emphasizes routing different tasks to different models, arguing that lower-cost DeepSeek Flash models are suitable for mechanical agent steps while more capable models or delegated frontier agents can handle harder work, with DeepInfra’s flat pricing sometimes comparing favorably with DeepSeek’s peak and off-peak rates. A practical build demonstrated that the harness could research and create a functioning voice-note workflow at low token cost, while also revealing limitations in workspace-write isolation, which restricts writes but not reads, and in the agent’s ability to overlook important system-level consequences. Compared with opencode and Claude Code, DeepSeek Harness is positioned as less mature but more composable, making it most relevant to developers building agent infrastructure, seeking configurable audit trails, or orchestrating multiple agents, while users needing stable, terminal-first, production-ready tooling may prefer established alternatives.

Multi-turn reinforcement learning for agents is presented as primarily an inference-scaling challenge rather than a gradient-optimization problem, because long, variable-length trajectories require repeated model calls and replay increasingly large conversation contexts. The discussion explains that rollout costs grow roughly quadratically with the number of turns due to expanding prompts, while synchronous training batches can leave GPUs underutilized as they wait for slow rollouts. It recommends decoupling rollout generation from policy training, using asynchronous worker pools and hosted APIs for fixed-weight tasks such as environment simulation, reward judging, baseline evaluation, and tool responses, while reserving local infrastructure for the continually changing open-weight policy. It also advocates choosing different models for policy, simulation, judging, and performance baselines based on their distinct cost and capability requirements, applying trajectory-level rewards with explicit penalties for failures such as malformed tool calls or turn-limit exhaustion, and separately tracking reward components to identify reward hacking or judge failures. Cost examples illustrate that judge inputs and simulator outputs can dominate spending, while prefix caching can substantially reduce repeated-context costs. Finally, it advises teams to establish independent evaluations and verify that a base policy has meaningful initial success—suggested at roughly 20%—before committing to expensive RL training, since stronger models, improved scaffolding, or agent frameworks may offer cheaper gains.

A 2026 comparison of AI inference platforms argues that selection should be based on cost per completed task alongside time to first token, output throughput, concurrency behavior, and endpoint type rather than token prices or isolated speed claims. It distinguishes shared serverless endpoints, which favor flexible per-token billing but can introduce batching, queueing, and cold-start latency, from dedicated capacity, which offers more predictable performance but charges by time and requires sufficient sustained demand. Using Llama 3.3 70B Instruct as a common benchmark, the analysis identifies DeepInfra and OpenRouter as the lowest-cost options, Groq and SambaNova as leading measured serverless throughput and responsiveness options, Cerebras as claiming exceptionally high throughput for a limited model catalog, Scaleway as an EU-residency-focused provider, Together AI as oriented toward combined fine-tuning and serving, Novita AI as a low-cost provider with dedicated deployment options, and Baseten as focused on custom, regulated deployments. It emphasizes that benchmark results vary substantially by model, quantization, prompt size, location, scheduling, and load, so organizations should test realistic workloads at expected concurrency and calculate costs using their own input-output token mix, caching, batch discounts, reasoning-token overhead, and duty cycle before choosing between serverless and dedicated infrastructure.