- Add 'playwright install chromium' step to Forge CI workflow
- Auto-detect default model from available API keys (ANTHROPIC_API_KEY,
OPENAI_API_KEY, GROQ_API_KEY) in direct_benchmark harness
- Prefer Claude > OpenAI > Groq, fallback to OpenAI if no keys found
Integrate standard AI agent benchmarks into the direct_benchmark infrastructure
using a plugin-based adapter pattern:
- Add BenchmarkAdapter base class with setup(), load_challenges(), and evaluate()
- Implement GAIAAdapter for the GAIA benchmark (requires HF token)
- Implement SWEBenchAdapter for SWE-bench (requires Docker)
- Implement AgentBenchAdapter for AgentBench multi-environment benchmark
- Extend HarnessConfig with benchmark options (--benchmark, --benchmark-split, etc.)
- Modify ParallelExecutor to use adapter's evaluate() for external benchmarks
- Fix runner to record finish step (was being skipped, breaking answer extraction)
- Add optional benchmarks dependency group with datasets and huggingface-hub
- Increase default benchmark timeout to 900s
Usage:
poetry run direct-benchmark run \
--benchmark agent-bench \
--benchmark-subset dbbench \
--strategies one_shot \
--models claude
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Update black version to match pre-commit hook (24.10.0) and reformat
all files with the new version.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add missing strategies (lats, multi_agent_debate) to PromptStrategyName
- Fix method override signatures for reasoning_effort parameter
- Fix Pydantic Field() overload issues with helper function
- Fix BeautifulSoup Tag type narrowing in web_fetch.py
- Fix Optional member access in playwright_browser.py and rewoo.py
- Convert hasattr patterns to getattr for proper type narrowing
- Add proper type casts for Literal types
- Fix file storage path type conversions
- Exclude legacy challenges/ from pyright checking
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Update .flake8 config to exclude workspace directories and ignore E203
- Fix import sorting (isort) across multiple files
- Fix code formatting (black) across multiple files
- Remove unused imports and fix line length issues (flake8)
- Fix f-strings without placeholders and unused variables
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add comprehensive sub-agent spawning infrastructure that enables prompt
strategies to coordinate multiple agents for advanced reasoning patterns.
New files:
- forge/agent/execution_context.py: ExecutionContext, ResourceBudget,
SubAgentHandle, and AgentFactory protocol for sub-agent lifecycle
- agent_factory/default_factory.py: DefaultAgentFactory implementation
- prompt_strategies/lats.py: Language Agent Tree Search using MCTS
with sub-agents for action expansion and evaluation
- prompt_strategies/multi_agent_debate.py: Multi-agent debate with
proposal, critique, and consensus phases
Key changes:
- BaseMultiStepPromptStrategy gains spawn_sub_agent(), run_sub_agent(),
spawn_and_run(), and run_parallel() methods
- Agent class accepts optional ExecutionContext and injects it into strategies
- Sub-agents enabled by default (enable_sub_agents=True)
- Resource limits: max_depth=5, max_sub_agents=25, max_cycles=25
All 7 strategies now available in benchmark:
one_shot, rewoo, plan_execute, reflexion, tree_of_thoughts, lats, multi_agent_debate
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Enable agents to execute shell commands during benchmarks by setting
execute_local_commands=True and using denylist mode to block dangerous
commands (rm, sudo, chmod, kill, etc.) while allowing safe operations.
Also adds ExecutePython challenge to test code execution capability.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Merge forge/, original_autogpt/, and direct_benchmark/ into a single Poetry
project to eliminate cross-project path dependency issues.
Changes:
- Create classic/pyproject.toml with merged dependencies from all three projects
- Remove individual pyproject.toml and poetry.lock files from subdirectories
- Update all CLAUDE.md files to reflect commands run from classic/ root
- Update all README.md files with new installation and usage instructions
All packages are now included via the packages directive:
- forge/forge (core agent framework)
- original_autogpt/autogpt (AutoGPT agent)
- direct_benchmark/direct_benchmark (benchmark harness)
CLI entry points preserved: autogpt, serve, direct-benchmark
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When a challenge times out but the agent's solution would have passed
evaluation, this is now clearly indicated:
- Completion blocks show "TIMEOUT (would have passed)" in yellow
- Recent completions panel shows hourglass icon + "would pass" suffix
- Summary table has new "Would Pass" column
- Final summary shows "+N would pass" count
- Success rate includes "would pass" challenges
The evaluator still runs on timed-out challenges to calculate the score,
but success remains False. This gives visibility into near-misses that
just needed more time.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously, the evaluator would run on all results including timed-out
challenges. If the agent happened to write a working solution before
timing out, evaluation would pass and override success=True, resulting
in contradictory output showing both PASS and "timed out".
Now we skip evaluation for timed-out challenges - they cannot pass.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The TicTacToe and other challenges use pytest-based test files for
evaluation. Without pytest installed in the benchmark virtualenv,
these evaluations were silently failing.
Root cause: test.py imports pytest but the package wasn't a dependency,
causing ModuleNotFoundError during evaluation subprocess.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Include config:challenge:attempt and timestamp in completion block
header for easier debugging and log correlation.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Remove selenium.py and test_selenium.py
- Add playwright_browser.py with WebPlaywrightComponent
- Update web component exports to use Playwright
- Update dependencies in pyproject.toml/poetry.lock
- Minor agent and reflexion strategy improvements
- Update CLAUDE.md documentation
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Previously cost was hardcoded to 0.0. Now extracts cumulative cost
from MultiProvider.get_incurred_cost() after each step execution.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Benchmarks now automatically save progress and resume from where they
left off. State is persisted to .benchmark_state.json in reports dir.
Features:
- Auto-resume: runs skip already-completed challenges
- --fresh: clear all state and start over
- --retry-failures: re-run only failed challenges
- --reset-strategy/model/challenge: selective resets
- `state show/clear/reset` subcommands for state management
- Config mismatch detection with auto-reset
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add --ci flag that disables Rich Live display while preserving
completion blocks. Auto-detects CI environment via CI env var or
non-TTY stdout. Prints progress every 10 completions for visibility.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Calculate max columns based on terminal width (up to 10)
- Reduced panel width from 35 to 30 chars to fit more
- Wider terminals can now show more parallel runs side-by-side
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Fixes:
- Use run_key (config:challenge) instead of just config_name for tracking
active runs - allows multiple challenges from same config to run in parallel
- Add asyncio.sleep(0) yields to let multiple tasks acquire semaphore
and start before any proceed with work
- Always print completion blocks (not just failures) for visibility
This should properly show 8/8 active runs when running with --parallel 8.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
UI improvements:
- Multi-column layout: each active config gets its own panel showing
challenge name and step history (last 6 steps with status)
- Copy-paste completion blocks: when a challenge finishes (especially
failures), prints a detailed block with all steps for easy debugging
- Configurable logging: suppresses noisy LLM provider warnings unless
--debug flag is set
- Pass debug flag through harness to UI
Example active runs panel:
┌─ one_shot/claude ─┬─ rewoo/claude ────┐
│ ReadFile │ WriteFile │
│ ✓ #1 read_file │ ✓ #1 think │
│ ✓ #2 write_file │ ✓ #2 plan │
│ ● step 3: ... │ ● step 3: ... │
└───────────────────┴───────────────────┘
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add step callback to AgentRunner for real-time step logging
- BenchmarkUI now shows:
- Active runs with current step info
- Recent steps panel with colored config prefixes
- Proper Live display refresh (implements __rich_console__)
- Each config gets a distinct color for easy identification
- Verbose mode prints step logs immediately with config prefix
- Fix Live display not updating (pass UI object, not rendered content)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Remove old benchmark/ folder with agbenchmark framework
- Move challenges to direct_benchmark/challenges/
- Move analysis tools (analyze_reports.py, analyze_failures.py) to direct_benchmark/
- Move challenges_already_beaten.json to direct_benchmark/
- Update CI workflow to use direct_benchmark
- Update CLAUDE.md files with new benchmarking instructions
- Add benchmarking section to original_autogpt/CLAUDE.md
The direct_benchmark harness directly instantiates agents without HTTP
server overhead, enabling parallel execution with asyncio semaphore.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
