V0.13.0

We are pleased to announce AmritaCore v0.13.0! This release introduces a native step‑loop architecture for the built‑in ReAct strategy, fundamentally improving the way agents decompose and execute complex tasks. Alongside this, we have completely restructured the documentation to better guide you from your first agent to deep internals.

---

✨ Major Features

Native Step Loop for ReAct Strategy

The built‑in ReActAgentStrategy now runs on a native instruction‑driven step loop, powered by AmritaSense’s NATIVE_WHILE and NATIVE_DO instructions.

• Task Decomposition – The LLM decides whether to break a task into a semantic DAG (Directed Acyclic Graph).
• Step‑by‑Step Execution – Each DAG node becomes a Step (intro → execute → leave).
• Stall Detection – Repeated identical tool calls within a Step trigger a "give‑up" prompt and stop the loop, preventing token waste.
• Lifecycle Events – New mutable events (agent.step_intro, agent.step_leave, agent.step_iteration, agent.tool_call, agent.tool_return) allow fine‑grained control and observability.
• update_step Tool – Agents can revise the plan mid‑run (replan, add/remove steps, mark done).
• Peer Message Injection – Use send_to_producer() to push messages from the consumer side; they are drained at Step boundaries and injected into the agent context.

This architecture makes agent execution more predictable, observable, and efficient—especially for multi‑step tasks.

Documentation Overhaul

We have completely re‑organised the documentation to mirror the natural development journey:

• Getting Started – Minimal and basic examples to run your first agent.
• Tutorials – Step‑by‑step guides for tools, streaming, events, and memory.
• Concepts – Deep dives into ChatObject, Agent Strategy, Data Backend, and the new Step Loop.
• Agent Engineering – Practical craft: prompt engineering, Jinja2 templates, custom strategies, and troubleshooting.
• Advanced – Workflow engine internals, suspend/resume, and the step loop deep‑dive.
• Extensions & Integration – Adapters, MCP servers, custom tokenizers, and advanced tool patterns.

The new structure helps you find exactly what you need at every stage—from first run to deep internals.

🔧 Improvements & Fixes

• Thinking‑mode round‑trip – Assistant messages now correctly carry reasoning_content back to providers (fixes HTTP 400 errors on DeepSeek thinking models).
• HybridReActAgentStrategy (deprecated) – Maintained with fixes, but will be removed in v0.14.0. Please migrate to ReActAgentStrategy.
• Event filtering – The thinking filter no longer mutates live message objects, ensuring reasoning content is preserved for subsequent requests.
• Stall detection – Now runs inside the iteration loop (after_iteration), so a stuck agent stops burning tokens immediately.
• Better error messages – Empty responses from decomposition/summary now include the provider’s request ID for easier debugging.

📚 Documentation Highlights

New pages added:

• AmritaSense Overview – The execution substrate explained.
• The Step Loop – Full anatomy of the new native step loop.
• Troubleshooting – Common failure modes and fixes.
• Custom Strategy – Step‑by‑step guide to writing your own strategy.
• MCP Servers – Simplified MCP integration guide.

Also updated: configuration, event system, data backend, and security sections.

⚠️ Deprecations & Breaking Changes

• HybridReActAgentStrategy – Deprecated in v0.13.0 and scheduled for removal in v0.14.0. Use ReActAgentStrategy instead.
• StateContext – Marked as legacy; use DI contexts (_di_memory, _di_ability, etc.) directly in new code.
• chat_object on StrategyContext – No longer deprecated; it remains the lifecycle‑manager handle. Prefer DI resource fields when available.
• agent.step_intro/leave/iteration events – These replace the old SINGLE_STRATEGY_CALL and REACT_COUNTER control flow for the built‑in ReAct strategy. If you have custom matchers that relied on the old loop internals, please review the new Step Events documentation.

🧹 Dependency Updates

• amrita-sense upgraded to >=0.6.0 (new NATIVE_WHILE/NATIVE_DO support).
• aiohttp → 3.14.3
• openai → 2.50.0
• cryptography → 50.0.0
• Plus several dev‑dependency updates (coverage, ruff, etc.).

PRs

• Build(deps): Bump postcss from 8.5.15 to 8.5.23 in /docs by @dependabot[bot] in #132
• Build(deps): Bump amrita-sense from 0.5.0 to 0.5.1 by @dependabot[bot] in #136
• Build(deps): Bump pytz from 2026.2 to 2026.3.post1 by @dependabot[bot] in #134
• Build(deps): Bump openai from 2.47.0 to 2.50.0 by @dependabot[bot] in #137
• Restructure docs and update API/security reference by @JohnRichard4096 in #138
• feat: implement native step loop and restructure docs for v0.13.0 by @JohnRichard4096 in #144
• Build(deps): Bump undici from 7.28.0 to 7.29.0 in /docs by @dependabot[bot] in #139
• Build(deps): Bump aiohttp from 3.14.2 to 3.14.3 by @dependabot[bot] in #133

---

Thank you to everyone who contributed to this release!

Try it today!

pip install amrita-core==0.13.0

Happy building 🚀

Full Changelog: 0.12.7...0.13.0