deepagents

Changes since deepagents==0.4.8 release(deepagents): 0.4.9 (#1773) feat(sdk): add prompt guidance for large tool results (#1763)

---

Thanks to our contributors: @sydney-runkle

Changes since deepagents==0.4.7 release(deepagents): 0.4.8 (#1764) test(sdk): bump langsmith-sdk for fixes experiments (#1752) feat(sdk): add factory function for summarization tool middleware (#1749) feat(sdk): add solve_rate metric to evals reporter (#1711) feat(cli): --acp mode to run CLI agent as ACP server (#1297) test(sdk): add metadata for experiments (#1652) test(sdk): reuse model fixture in summarization evals (#1697) feat(sdk): add baseten-hosted models to evals workflow (#1682)

Changes since deepagents==0.4.6 release(sdk): bump deepagents to 0.4.7 (#1688) fix(deepagents): restore full routed path in write/edit results (#1681) feat(infra): add fireworks as eval provider (#1684)

---

Thanks to our contributors: @naga-k

Changes since deepagents==0.4.5 release(sdk): bump deepagents to 0.4.6 (#1677) chore: bump lock file (#1674) fix(sdk): extract text from all content blocks for large tool result eviction (#1673) fix(sdk): enforce path boundaries in CompositeBackend routing (#1659) refactor(sdk): split eval assertions into correctness and efficiency tiers (#1670) test(sdk): remove an ambiguous eval test (#1661) chore(sdk): add tool usage evals (#1637) test(deepagents): fix one more file operation test (#1653) test(sdk): update eval test to be a bit less ambiguous (#1643) docs(sdk): middleware module docstring (#1620) chore(sdk): expand eval matrix and fix model name access for groq (#1623)

---

Thanks to our contributors: @dotuananh0712, @bracesproul, @eyurtsev, @mdrxy

Changes since deepagents==0.4.4 release(deepagents): 0.4.5 (#1622) chore(sdk): expand eval model matrix to 28 models across 7 providers (#1617) chore: remove unnecessary @pytest.mark.asyncio decorators (#1616) chore(sdk): update evals to use chat model instance (#1614) fix(sdk): make evals more robust across model output variations (#1608) fix(infra): streamline evals workflow and summary (#1602) docs(sdk): improve docstrings around summarization conditions (#1597) chore(sdk): add json report to evals (#1594) chore(sdk,cli): switch claude-sonnet-4-5-20250929 -> claude-sonnet-4-6 (#1590) feat(cli): /compact command (#1579) chore: standardize makefile type targets/ add type CI check (#1581) ci(sdk): skip some evals (#1580) docs(sdk): rename 'prompt' to 'system_prompt' in subagent keys (#1578) docs(sdk): docstring fixes for summarization.py (#1568) chore(sdk): remove unused base_path alias in skills.py (#1543) chore(deps): bump the pip-dependencies group across 8 directories with 17 updates (#1564) fix(sdk,cli,deps): support timeout=0 for sandbox execute() (#1558) fix: Potential fixes for 2 code quality findings (#1534) fix: Variable defined multiple times (#1536) feat(cli,sdk): compaction hook (#1420) chore(deps): bump the pip-dependencies group across 8 directories with 19 updates (#1491)

---

Thanks to our contributors: @AdemBoukhris457, @AlexPapadakis, @eyurtsev

Changes since deepagents==0.4.3 release(deepagents): 0.4.4 (#1519) fix(sdk): use consistent path stripping in CompositeBackend grep and glob (#1504) fix(sdk): add timeout to glob in FilesystemMiddleware (#1508) fix(sdk): validate tool_call_id before subagent invocation (#1506) fix(sdk): bundle read template params into single base64 payload (#1500) fix(deepagents): preserve ToolMessage artifact when evicting large outputs (#1482) fix(sdk): fix glob matching with ** in _python_search fallback (#1350) fix(sdk): sanitize paths in BaseSandbox ls_info and read templates (#1484) docs(sdk): fix SkillsMiddleware.before_agent docstring to reflect caching behavior (#1495) perf(sdk): parallelize summarization offload and LLM summary generation (#1488) chore(sdk): add coverage to Makefile (#1493) perf(sdk): batch memory source loading into a single download_files call (#1490)

---

Thanks to our contributors: @dotuananh0712, @AdemBoukhris457, @OiPunk

Changes since deepagents==0.4.2 release(deepagents): 0.4.3 (#1465) fix(sdk): guard timeout forwarding for older backend packages (#1461) fix(sdk): correct imports and mock inheritance in filesystem middleware tests (#1464) fix(sdk): warn on virtual mode (#1462) chore(sdk): add more eval test cases (#1452) test(deepagents): add evaluation tests for skills (#1429) fix(deepagents): Fix Memory prompt guidelines that could lead to wrong actions (#1448) test(sdk): add smoke test for system prompt with memory and skills (#1430) chore(deepagents): bump min version for langchain-anthropic (#1431) feat(deepagents): add evaluation tests for agent memory (#1415) chore(sdk): adapt integration tests to evals framework (#1426) feat(sdk): add per-command timeout override to execute() (#1154) chore: auto-generate makefile help from target annotations (#1417) chore(sdk): add more evals (#1404) chore(sdk): minor updates to system message (#1367) chore(sdk): add model parameterization to evals (#1408) chore: revert SDK version following yank (#1407) chore(sdk): add simple eval scaffolding (#1400) refactor(sdk): add logging to silent path skips, rename validate_path (#1370) fix(cli,sdk): harden path hardening (#918) chore(sdk): add smoke tests for system message (#1363) feat(sdk): enable type checking in deepagents and resolve most linting issues (#991) chore(sdk): move some integration tests into unit tests (#1330) revert(sdk): reimplement skill loading via skill tool calling (#1328) fix(sdk): missing _validate_path() on Glob tool's path parameter allows directory traversal (#1320) fix: Path Traversal via Unsanitized Glob Pattern in glob_info() Bypasses Virtual Mode Sandbox (#1322) feat(sdk): add base agent prompt (#1314)

---

Thanks to our contributors: @MuriloAndre2000, @vtrivedy, @sydney-runkle, @app/corridor-security, @app/copilot-swe-agent, @ccurme, @eyurtsev, @mdrxy

Changes since deepagents==0.4.1 release(sdk): 0.4.2 (#1312) feat(sdk): reimplement skill loading via skill tool calling (#1306)

---

Thanks to our contributors: @sydney-runkle

Changes since deepagents==0.4.0 feat(sdk): return native image content blocks from read_file (#1163) chore(deps): bump the pip-dependencies group across 8 directories with 9 updates (#1252) refactor(sdk): use public SummarizationMiddleware middleware (#1241) chore: rename master to main (#984)

---

Thanks to our contributors: @app/dependabot, @mdrxy

Changes since deepagents==0.3.12

New integration packages for pluggable sandboxes: langchain-modal, langchain-daytona, and langchain-runloop. See sandboxes guide and example data analysis tutorial.

Notable changes from #1234

• Refactor the SDK's summarization middleware.
  • Summarization now happens in the model node via wrap_model_call events. Due to this we retain the full message history in the graph state.
  • System messages and tool definitions are accurately counted toward summarization triggers.
  • Add exception handling for context overflows: ChatOpenAI and ChatAnthropic now raise a standard LangChain exception when input exceeds a model's token limit. We catch this error and trigger summarization.
• When model strings in create_deep_agent are prefixed with 'openai:', we default to the Responses API.

chore(sdk): add support for , in allowed-tools, fail gently for bullet list (#1235) fix(sdk): relax skills parsing for allowed tools (#1232) chore(sdk)!: move sandbox provider back to cli (#1226) refactor(sdk): move function-level test imports out to top (#1168) fix(sdk): harden skills metadata parsing, improve spec compliance (#1189) style: update brand names (#1180)

---

Thanks to our contributors: @sydney-runkle, @eyurtsev, @mdrxy

Changes since deepagents==0.3.11

• feat(sdk): refactor namespace gen for StoreBackend (#1169)
• fix(sdk): add explicit NotImplementedError to protocol method bodies (#1152)
• refactor(cli): reorder property -> method order in SandboxBackendProtocol (#1153)
• feat(sdk): add sandbox provider exceptions (#1108)

---

Thanks to our contributors: @eyurtsev, @mdrxy

Changes since deepagents==0.3.10

• feat(sdk): subagent skills, middleware refactor (#1041)
• fix(sdk): add newlines in tool result message for large outputs (#1053)

---

Thanks to our contributors: @sydney-runkle

Changes since deepagents==0.3.9

• fix(sdk): grep should perform literal search instead of regex (#975)
• feat(sdk): sandbox provider interface (#900)
• fix(sdk): handle exact file paths in grep and glob operations (#1017)
• fix: add py.typed to deepagents (#1024)
• fix(sdk): return error message instead of raising ValueError for invalid paths (#994)
• fix(deepagents): prevent state merge error in parallel sub-agents (#954)
• chore: enrich pyproject.toml files (#996)
• chore(sdk,cli): add VERSION (#983)
• feat(sdk): add LocalShellBackend (#930)
• fix(sdk): string interpolation in BaseSandbox write/edit tools (#955)

---

Thanks to our contributors: @vtrivedy, @sydney-runkle, @sandeepyadav1478, @Yo-sure, @trindadetiago, @eyurtsev, @mdrxy

Changes since deepagents==0.3.8

release(deepagents): 0.3.9 (#944) fix(sdk): summarization offloading (#920) refactor: filesystem middleware (#928) chore(deps): bump the uv group across 3 directories with 6 updates (#905) test(sdk): add unit test for handling large single-line file bug (#813) feat(deepagents): truncate middle of content instead of end on tool response eviction (#901) feat(deepagents): truncate tool messages from read_file (#897) fix(sdk,acp,harbor): don't allow relative imports (#882) fix(sdk): handle large payloads and headless container execution (#872) chore: update Makefile targets, add pre-commit hooks (#790)

Changes since deepagents==0.3.7

release(sdk): bump version (#879) fix(sdk): make sure that tool truncation applies to execute (#547) test(sdk): use unique thread_id for summarization test configuration (#871)

Changes since deepagents==0.3.6

fix(sdk): don't dedent summarization prompt (#870) feat(deepagents): truncate old write / edit calls in message history (#806) release: patch release 0.3.7 (#869) chore(deepagents): add end to end tests to confirm file reducer working properly in state backend (#754) chore: improve filesystem and subagents tool descriptions (#807) docs(SDK): clarify usage of file system backend (#850) nit: standardize naming (#849) docs(sdk): FilesystemBackend refs fixes (#791) feat: show allowed tools for each skill in skill list (#837) fix(sdk): FilesystemMiddleware forward 'name' attribute in large_tool_result from the original tool msg (#825) docs(sdk): docstring formatting nits (#824) feat: summarization offloading (#742) Bump version to 0.3.7a1 (#817) Add Async ops to Store Backend (#816) chore(deepagents): add tests for grep in end to end tests (#805) chore(deepagents): bump langchain in lock file (#800) chore(deps): bump the uv group across 5 directories with 1 update (#811) fix(deepagents): respect continuation markers when reading files (#809) fix(infra): exclude build/ from typechecking (#808) feat: support SystemMessage for parity w/ create_agent (#803) chore(deepagnets): end to end tests for agent writing editing files (#804) fix(sdk): BaseSandbox.ls_info() to return absolute paths (#797) fix(deepagents): truncate lines on read (#784) chore(deps): bump the uv group across 3 directories with 3 updates (#796) fix: refinements for test_summarization (#786) docs: fix old URLs (#787) docs: add testing readme (#788) fix: added error catching for file operations without permissions (#734) docs(deepagents): update subagent spec (#785) chore(deepagents): add mini eval for summarization (#751) docs(sdk): improve FileSystemBackend ref docs (#783)

Initial release

Bump version to 0.3.7a1 (#817) Add Async ops to Store Backend (#816) chore(deepagents): add tests for grep in end to end tests (#805) chore(deepagents): bump langchain in lock file (#800) chore(deps): bump the uv group across 5 directories with 1 update (#811) fix(deepagents): respect continuation markers when reading files (#809) fix(infra): exclude build/ from typechecking (#808) feat: support SystemMessage for parity w/ create_agent (#803) chore(deepagnets): end to end tests for agent writing editing files (#804) fix(sdk): BaseSandbox.ls_info() to return absolute paths (#797) fix(deepagents): truncate lines on read (#784) chore(deps): bump the uv group across 3 directories with 3 updates (#796) fix: refinements for test_summarization (#786) docs: fix old URLs (#787) docs: add testing readme (#788) fix: added error catching for file operations without permissions (#734) docs(deepagents): update subagent spec (#785) chore(deepagents): add mini eval for summarization (#751) docs(sdk): improve FileSystemBackend ref docs (#783) chore: add tests for config, context, metadata propagation (#776) fix: import rules (#763) fix(deepagents): throw a clear exception message when 'messages' key missing from the output of the subagent (#678) release: deepagents 0.3.6 (#752) chore: automatically sort imports (#740) fix: no f string needed (#750) feat: add agent name when creating subagents (#735) feat(cli): inject local context into system prompt via LocalContextMiddleware chore: update twitter URL (#727) chore: CR on docstrings/readme (#722) minor release(deepagents): bump version to 0.3.5 (#695) Port SDK Memory to CLI (#691) chore(deepagents): turn on strict linting and add ratchet (#690) chore(deepagents): add tests for todos middleware (#667) fix(deepagents): fix content block handling in filesystem middleware tool result eviction (#674) Minor release memory bump (#689) feat: update memory prompt to be more barebones (#685) improvements to README, reference docs (#688) update version bounds (#687) Deepagents release minor (#683) fix(deepagents): make memory permissive if AGENTS.md not found (#682) release(deepagents): release 0.3.2 (#680) chore(deepagents): make memory strict if AGENTS.md not found (#673) fix(deepagents): better async support in skills, and propagate runnable config (#672) chore: remove older integration tests for memory (#670) feature(deepagents): add memory (#646) feat(deepagents): add skills to sdk (#591) chore(deepagent): add docs to composite backend (#666) chore(deepagents): add more test coverage to composite backend (#660) fix(deepagents): composite backend grep implementation (#659) Fix CVE-2025-68664 (#636) make work with model string (#626) release(deepagents): 0.3.1 (#608) docs: fix documentation issues (#513) fix(deepagents): strip trailing whitespace from subagent messages to prevent Anthropic API errors (#586) fix(deepagents): Pass through runtime config to subagents (#602) chore(deepagents): test write todos from sub-agents (#605) fix(deepagents): exclude structured response from state update (#603) feat: add ability to paste images in input (#555) chore: bump minor version (#549) feature(deepagents)!: add async implementations for backend, sandbox, filesystem (#504) fix: fix docstring formatting errors in backends (#514) chore(cli): add comprehensive testing for sandbox operations (#501) feat(deepagents): use context window fractions for SummarizationMiddleware when available (#484) chore(deepagents): add end to end test (#481) release(deepagents, cli) (#477) return strings in tools (#474) truncate glob (#473) fix(deepagents): fix type for file info (#471) add file upload and download apis (#381) fix(deepagents-cli): fix linting (broke CI) (#457) fix(middleware): prevent silent corruption from Windows absolute paths (Issue #427) (#454) Remove unnecessary dependencies from deepagents module (#429) fix: use request.override instead of direct attribute overrides (#431) Update README.md (#422) release: deepagents 0.2.9, cli 0.0.9 (#411) patch: remove resumable shell middleware (#410) release (#388) chore: finish migration into monorepo structure (#383) feat: expose id on base abstraction (#382) feat: sandbox protocol (#319) chore: move BACKEND_TYPES to correct place (#351) fix: Don't error when "subagent" does not exist. And add test for truncation logic (#343) minor update (#302) chore: inherit from protocol, update types (#291) Truncate content sample (#287) chore(deepagents): move protocol types (#286) move deepagents to libs/deepagents (#257)

Changes since deepagents==0.3.5

chore: add tests for config, context, metadata propagation (#776) fix: import rules (#763) fix(deepagents): throw a clear exception message when 'messages' key missing from the output of the subagent (#678) release: deepagents 0.3.6 (#752) chore: automatically sort imports (#740) fix: no f string needed (#750) feat: add agent name when creating subagents (#735) feat(cli): inject local context into system prompt via LocalContextMiddleware chore: update twitter URL (#727) chore: CR on docstrings/readme (#722)

Changes since deepagents==0.3.4

minor release(deepagents): bump version to 0.3.5 (#695) Port SDK Memory to CLI (#691) chore(deepagents): turn on strict linting and add ratchet (#690) chore(deepagents): add tests for todos middleware (#667) fix(deepagents): fix content block handling in filesystem middleware tool result eviction (#674)

Changes since deepagents==0.3.3

Minor release memory bump (#689) feat: update memory prompt to be more barebones (#685) improvements to README, reference docs (#688) update version bounds (#687)

Changes since deepagents==0.3.2

Deepagents release minor (#683) fix(deepagents): make memory permissive if AGENTS.md not found (#682)

Changes since deepagents==0.3.1

release(deepagents): release 0.3.2 (#680) chore(deepagents): make memory strict if AGENTS.md not found (#673) fix(deepagents): better async support in skills, and propagate runnable config (#672) chore: remove older integration tests for memory (#670) feature(deepagents): add memory (#646) feat(deepagents): add skills to sdk (#591) chore(deepagent): add docs to composite backend (#666) chore(deepagents): add more test coverage to composite backend (#660) fix(deepagents): composite backend grep implementation (#659) Fix CVE-2025-68664 (#636) make work with model string (#626)

Changes since deepagents==0.3.0

release(deepagents): 0.3.1 (#608) docs: fix documentation issues (#513) fix(deepagents): strip trailing whitespace from subagent messages to prevent Anthropic API errors (#586) fix(deepagents): Pass through runtime config to subagents (#602) chore(deepagents): test write todos from sub-agents (#605) fix(deepagents): exclude structured response from state update (#603) feat: add ability to paste images in input (#555)

Changes since deepagents==0.2.8

chore: bump minor version (#549) feature(deepagents)!: add async implementations for backend, sandbox, filesystem (#504) fix: fix docstring formatting errors in backends (#514) chore(cli): add comprehensive testing for sandbox operations (#501) feat(deepagents): use context window fractions for SummarizationMiddleware when available (#484) chore(deepagents): add end to end test (#481)

Changes since deepagents==0.2.7

release(deepagents, cli) (#477) return strings in tools (#474) truncate glob (#473) fix(deepagents): fix type for file info (#471) add file upload and download apis (#381) fix(deepagents-cli): fix linting (broke CI) (#457) fix(middleware): prevent silent corruption from Windows absolute paths (Issue #427) (#454) Remove unnecessary dependencies from deepagents module (#429) fix: use request.override instead of direct attribute overrides (#431) Update README.md (#422)

Changes since deepagents==0.2.6

release: deepagents 0.2.9, cli 0.0.9 (#411) patch: remove resumable shell middleware (#410)

Changes since deepagents==0.2.5

release (#388) chore: finish migration into monorepo structure (#383) feat: expose id on base abstraction (#382) feat: sandbox protocol (#319) chore: move BACKEND_TYPES to correct place (#351) fix: Don't error when "subagent" does not exist. And add test for truncation logic (#343) minor update (#302)

Changes since deepagents==0.2.4

ci: disable running prerelease tests temporarily (#296) quick packaging fix (#295) ci: add release workflow (wip) (#289) Patch release 0.2.5 (#294) ci: enable format lint (#292) chore: inherit from protocol, update types (#291) Truncate content sample (#287) chore: gitignore (#288) chore(deepagents): move protocol types (#286) fix (cli): autocomplete for @ and / commands through directories, bash mode in TUI (#278) Add Dependabot configuration file (#285) chore: delete accidental file (#282) handle hilt in a way that makes tracing better (#277) simplify hilt (#276) move deepagents to libs/deepagents (#257) ci: no need for secrets atm (#256) chore: add ci (#254) chore: linting pass (#253)

Oct 31, 2025 2 files

Oct 30, 2025 2 files

Oct 30, 2025 2 files

Oct 30, 2025 2 files

Pre-release

Oct 29, 2025 2 files

Pre-release

Oct 29, 2025 2 files

Oct 28, 2025 2 files

Pre-release

Oct 27, 2025 2 files

Pre-release

Oct 27, 2025 2 files

Oct 23, 2025 2 files

Oct 21, 2025 2 files

Oct 20, 2025 2 files

Oct 18, 2025 2 files

Oct 17, 2025 2 files

Pre-release

Oct 17, 2025 2 files

Pre-release

Oct 16, 2025 2 files

Pre-release

Oct 8, 2025 2 files

Oct 7, 2025 2 files

Pre-release

Sep 30, 2025 2 files

Sep 29, 2025 2 files

Sep 26, 2025 2 files

Sep 25, 2025 2 files

Sep 24, 2025 2 files

Sep 24, 2025 2 files

Pre-release

Sep 23, 2025 2 files

Pre-release

Sep 19, 2025 2 files

Pre-release

Sep 17, 2025 2 files

Sep 1, 2025 2 files

Pre-release

Sep 4, 2025 2 files

Pre-release

Aug 31, 2025 2 files

Pre-release