Agents that keep what works: FlowEvo turns workflows into reusable skills
FlowEvo compiles successful action sequences into callable skills at runtime and stores them for later use — delivering 85.6% on ALFWorld while using about one-third the tokens. A new unlearning benchmark and an Agents SDK update round out today’s push toward sturdier, safer agents.
One-Line Summary
Agents learn during use by saving winning steps, while safety-focused unlearning is tested at the concept level and tooling for multi-agent workflows hardens.
Research Papers
FlowEvo: Agents that save and reuse their own skills
FlowEvo is a training-free way for an agent to keep and reuse what works: it compiles successful action sequences into callable skills at runtime and stores them in a persistent bank. Built on the same backbone Large Language Model (LLM) — GPT-4o-mini — across tests, FlowEvo achieves the highest accuracy among eight baselines on ALFWorld, HumanEval, MBPP, GSM8K, and MATH-500, while using roughly one third as many tokens. On ALFWorld it scores 85.6%, a 26.4-point gain over the strongest baseline. 1
Instead of treating each run as disposable, FlowEvo turns any effective workflow into an executable skill, records it in a skill bank, and retrieves it later either to run directly or to guide construction of new workflows. The system tracks each skill’s downstream utility and suppresses ones that cause negative transfer, creating a co-evolution loop between the agent’s workflows and its growing library — all without additional training. 1
Across 10 base models ranging from 7B to 671B parameters, FlowEvo outperforms the prior ExpeL approach in 49 of 50 model–dataset comparisons. The authors release code and report consistent gains across coding, math, and embodied tasks using the shared GPT-4o-mini backbone. 1
That design echoes a broader shift described as Agent-Centric Interactive World Proxies: building feedback engines that let agents improve at inference time, optimize during training, and co-evolve with their proxies. This framing argues for grounded, controllable simulators (from execution to memory and skill guidance) to reduce risky real-world trials while accelerating agent learning. 2
ConceptGuard: Measuring context-sensitive unlearning in LLMs
ConceptGuard is a benchmark that tests whether a model can forget harmful uses of a concept while preserving its benign uses. Unlike prior setups that separate unrelated facts into “forget” and “retain” buckets, ConceptGuard builds complementary sets around dual-use concepts and evaluates intent-sensitive behavior — a more practical measure of unlearning for Large Language Models (LLMs). 3
The authors find that current unlearning methods struggle under this setting, showing weak contextual separation and drops in ROUGE and concept-level metrics, with strong trade-offs between forgetting and retaining utility. The dataset is public, inviting methods that control concepts, not just facts, to meet real-world safety needs. 3
Open Source & Repos
OpenAI Agents SDK (Python) simplifies multi-agent workflows
OpenAI’s Agents SDK (Software Development Kit) for Python is a lightweight framework for building multi-agent workflows across providers. It supports the OpenAI Responses and Chat Completions Application Programming Interfaces (APIs) and more than 100 other Large Language Models (LLMs), and includes core concepts such as agents, sandbox agents for long-running work, and realtime agents. 4
Release v0.22.0 (Aug 19, 2026) adds substantial runtime hardening and tightens an existing provider configuration contract; apps that combine an explicit openai_client with organization or project must move those values to the AsyncOpenAI client. This points to a maturing focus on reliability and clearer configuration in multi-agent orchestration. 4
Why It Matters
The center of gravity in agents is shifting from one-off prompt chains to persistent skill libraries that compound over time — FlowEvo’s co-evolving bank shows accuracy gains and lower token costs without extra training. 1
Tooling is also maturing: a hardened, provider-agnostic Agents SDK clarifies configuration and supports multi-agent orchestration, making it easier to operationalize research ideas quickly. 4
This Week, Try It
- OpenAI Agents SDK quickstart: Install the Python package and run a multi-agent example from the README. https://github.com/openai/openai-agents-python
- Read ConceptGuard’s examples: Skim the arXiv paper’s task design to see how intent-sensitive unlearning is tested. https://arxiv.org/abs/2608.20338v1
Comments (0)