DeepSeek Harness Open-Source Coding Agent Framework Offers Four Distinct Operating Modes

·1 views

DeepSeek has released Harness, an open-source, MIT-licensed framework for building coding agents from modular, replaceable components including models, tools, sessions, and sandboxes. The project is labeled a Developer Preview and follows the design principle that an agent equals a model plus a harness, where the model handles reasoning and the harness manages context and execution loops. Built on the Cordis plugin system, Harness supports four modes — Standard, PTC, Minimal, and Creation — each suited to different use cases ranging from full coding-agent workflows to diagnostic baselines and custom preset design. Developers are advised to begin with a small, read-only task in an isolated test folder using Standard mode before attempting any edits or production deployment. The framework also cautions that each added plugin can alter the data, permissions, and external services accessible to the agent, placing configuration responsibility on the developer.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How to Measure Real Savings from Prompt Caching Across AI Providers

Prompt caching reduces reprocessing costs by reusing unchanged portions of an AI model's input, such as system prompts or reference documents, across multiple requests. To accurately measure its impact, developers should run a controlled three-request experiment: one to create the cache, one to trigger a cache hit, and one to force a cache miss by altering the prefix. Key variables like model ID, token thresholds, time-to-live, and pricing must be recorded for results to be meaningful. OpenAI applies caching automatically to eligible prefixes, while Anthropic requires explicit cache boundary markers via cache_control, and the two systems use different usage field names and cost structures. Developers can verify their own cache behavior and token consumption using provider dashboards, API references, and tools like BetterToken alongside official documentation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Handle API 429 Errors: Rate Limits, Retry Logic, and Safe Strategies

An HTTP 429 error indicates a server has rejected a request due to rate or usage limits, but the exact cause can vary widely across providers. RFC 6585 defines the status as 'too many requests,' yet it does not standardize how limits are counted or how long a client should wait before retrying. Responses may include a Retry-After header, though this is not guaranteed, and limits can apply at the level of a single resource, API key, project, or entire organization. Safe retry strategies involve reading the full response body and headers, respecting any Retry-After value as a minimum delay, and adding random jitter to avoid synchronized retry storms. For requests with side effects, developers must first confirm whether the original request completed before issuing another attempt, and repeated retries without addressing the root cause can worsen the problem.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Diagnose and Fix API Timeouts at the Right Network Layer

API timeouts do not always mean a model or service is down — they can originate at multiple points in the request chain, including DNS, TCP/TLS, proxy, gateway, or upstream model layers. Developers are advised to identify the exact failure point before adjusting any timeout settings, rather than broadly increasing all values at once. Key diagnostic steps include recording start and end times, exception classes, HTTP status codes, and request identifiers returned by APIs such as Claude's request-id. Different timeout types — connect, read, idle, and overall deadline — each carry distinct signals and require separate fixes. For long-running requests, providers like Anthropic recommend using streaming or batch processing to reduce the risk of timeouts exhausting client resources.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Fix 'Context Length Exceeded' Errors in LLM API Requests

When an LLM API returns a 'context length exceeded' error, the fix requires understanding that total context includes system instructions, conversation history, messages, files, tool definitions, tool results, and the output budget combined. Developers should first identify the model and measure each component's token usage before making any reductions. Safe trimming steps include removing duplicate system rules, irrelevant conversation history, redundant RAG chunks, unused tool definitions, and oversized tool results. Critical facts, accepted decisions, and mandatory constraints should be preserved, with large data stored outside the prompt rather than discarded. After resizing the request, both the absence of the error and the completeness of the response should be verified to ensure nothing essential was lost.

0 comments Read more at DEV Community