Download Latest Version v0.13.0 source code.tar.gz (7.0 MB)
Email in envelope

Get an email when there's a new version of AG2

Home / v0.12.2
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-05-01 6.2 kB
0
v0.12.2 source code.tar.gz 2026-05-01 6.6 MB
1 1 weekly downloads
v0.12.2 source code.zip 2026-05-01 8.1 MB
1 1 weekly downloads
Totals: 3 Items   14.7 MB 2

🛣️ Continuing the Path to v1.0

v0.12.2 lands the Agent Harness in Beta — the architecture that makes agents stateful by design.

📍 Release Roadmap 🧩 AG2 Playground

🏗️ The Agent Harness — Stateful Agents, by Design

A plain LLM loop is stateless and forgets the moment a conversation ends. The Harness turns it into a stateful agent that remembers across sessions, manages its own context window, organizes its own knowledge, and shares findings with peers — without rewriting your loop.

It does this by decomposing every agentic operation into four orthogonal stages, each a swappable Protocol you opt into:

  • 💾 Persistence – What the agent knows. Durable, cross-conversation.
  • 🧩 Assembly – What the agent sees. Composed per LLM call.
  • ⚡ Execution – The LLM call. Already exists in any loop.
  • 🔄 Post-Processing – Maintenance. Compaction · aggregation.

What this unlocks

  • 🕐 Long-running single agents – Hours without blowing the context window.
  • 🧠 Cross-session memory – "My name is Alice, working on project X" is still remembered tomorrow.
  • 🤝 Multi-agent knowledge sharing – Peers read each other's exposed paths via cheap reads, not expensive delegation round-trips.

What's in the box

  • 📁 KnowledgeStore – Agent-owned virtual filesystem with read/write/list/delete/exists. LLMs already speak filesystem — no new mental model. Five backends ship: MemoryKnowledgeStore, DiskKnowledgeStore, SqliteKnowledgeStore, RedisKnowledgeStore, LockedKnowledgeStore.
  • 🧩 Assembly Policies – Composable context shaping (not just a filter). Stack reduction and injection policies independently: ConversationPolicy, SlidingWindowPolicy, TokenBudgetPolicy, EpisodicMemoryPolicy, WorkingMemoryPolicy, AlertPolicy.
  • ✂️ Compaction – Constraint-respecting reduction of the live event stream. TailWindowCompact (zero LLM cost) or SummarizeCompact (1 LLM call per fire).
  • 📚 Aggregation – Builds long-term memory from short-term experience. ConversationSummaryAggregate writes per-session summaries; WorkingMemoryAggregate maintains persistent agent state.
  • 🚀 Bootstrap + SKILL.md – Self-describing knowledge stores. DefaultBootstrap writes /log/, /artifacts/, /memory/ with SKILL.md files so agents know where to look on first run.
  • 💰 Cost Visibility – Every harness LLM call emits a CompactionCompleted or AggregationCompleted event on the stream. Telemetry-ready, not buried in your bill.

Every feature is opt-in — a bare Agent still has zero harness middleware.

🔬 Other Beta Updates

  • 🔍 PerplexitySearchTool (Beta) – Now available in Beta and integrated into the search module.
  • 🔧 Normalized Parts on Builtin Tool Result Events – Tool result events now expose normalized parts.
  • 🔒 Server-Side Builtin Tool Calls – Now persisted in history.
  • ⚙️ Agent Subtasks Disabled by Default – Safer default behavior for Agent subtasks.

Beta Bug Fixes

  • 🔧 Beta: Gemini ToolCallEvent IDs – Generate unique IDs per ToolCallEvent when the provider omits fc.id.
  • 🔧 Beta: CadenceWatch Buffer – Drain buffer when callback outlasts max_wait.
  • 🔧 Beta: Redis Serialization – Moved serialization for Redis to the event serializer.
  • 🔧 Beta: Gemini Empty Text Parts – Drop empty text parts plus AG-UI guard.

What's Changed

Full Changelog: https://github.com/ag2ai/ag2/compare/v0.12.1...v0.12.2

Source: README.md, updated 2026-05-01