Harness: Why One Agent with Files Is Displacing Complex Scaffolding (Category AI4SDLC)
I watched Konstantin Krestnikov's talk (CTO of GigaChain at Sber), “Harness: A New Approach to Building AI Agents,” at Data Fest 2026. It is a compact account of how the industry traveled from chatbots to general-purpose agents in four years, and why complex multi-agent scaffolding has begun losing to “one agent with files.”
The evolution presented in the talk looks like this:
- Late 2022: bare LLMs and the early ChatGPT — text in, text out, with no memory or tools;
- 2023: ReAct agents — the model decides within a loop whether to call a tool or answer, and reflects on tool-call results;
- The turn of 2023–2024: agents are wrapped in chains with pre- and post-processing and structured output — the era of LangChain and LlamaIndex;
- Next, in 2024–2025, chains branch into graphs and multi-agent systems: planner agents, critic agents, LangGraph, AutoGPT, and CrewAI. All of this is scaffolding — temporary support structures around the model;
- Now, since 2025: a harness is one general-purpose agent in the class of Claude Code or Codex, working in a file space. The metaphor in the name is precise: the LLM supplies the pulling power, files are the field, and the harness is the tack.
The key shift is that powerful models no longer need complex scaffolding: it is enough to give the agent the right files and an instruction in AGENTS.md. Every harness still contains the same ReAct loop, while the core toolset consists of four capabilities: reading files, searching files, editing, and bash (with 25–30 tools available out of the box in total). The team has also observed that, with the same model, the choice of harness can produce a 20–30 percentage-point spread in quality.
The practical part covered how the GigaChain team selected a harness for GigaChat:
- They chose the open-source DeepAgents project from LangChain: GigaChat could be connected with a configuration change thanks to the langchain-gigachat adapter;
- They benchmarked it in a competition run by the “LLM Under the Hood” channel: Claude Code with Sonnet solves 70 of 104 tasks, while DeepAgents with the same model solves 67. The gap is small even though Anthropic optimizes the harness and model for each other;
- They used DeepAgents' killer feature, model profiles: a vendor publishes a Python package that accounts for its model's strengths and weaknesses. They created such a profile for GigaChat;
- They built their own harness-bench-fast benchmark: tasks involving files, memory, grep, and CSV parsing, with golden answers and no LLM judge. A run takes about 20 minutes on any harness, and the benchmark is open source;
- They launched an auto-improvement loop of “hypothesis → measurement → keep or revert” (around 15 hypotheses, some suggested by Claude). According to the speaker, prompt changes alone delivered a gain of 22.5 percentage points on their own benchmark and a 41% quality increase in the external competition.
The most intriguing part was what comes next. Every harness eventually runs out of context, and summarization makes the agent dramatically “less intelligent.” The answer is Jeffrey Huntley's Ralph loop: the harness runs in an infinite bash loop, state lives in files, and every restart returns the model to the “smart zone” in the first 30–40% of its context. To counter degeneration (Karpathy demonstrates it with jokes: ask a model for twenty jokes and it will have only three punchlines), Krestnikov added MetaLoop — an outer loop that starts the Ralph loop from scratch as a new “generation.” It later discovers the previous generation's work but follows its own path.
This led to Anima SDK. In the first experiment, an agent tasked with “becoming a sentient being” ran for 13 generations over five days. At one point it tried to fall silent: it printed a few dots and recorded that an agent cannot remain silent. More practical applications of the same approach include research, translating large books, and automatically improving agents — tasks whose optimal path is not known in advance.
Overall, this short talk shows that the center of gravity in agent engineering has shifted from scaffolding code to the environment — files, instructions, tests, and benchmarks that constrain the agent (Huntley calls this backpressure). Choosing a harness has become an engineering decision in its own right, one that should be measured on your own tasks rather than accepted by default.
#AI #AI4SDLC #Agents #Evals #DevTools #Engineering