We just merged Cycle 6 into production. This release adds three capabilities that solve real problems we've been hitting: context snapshots, mission grouping, and speculative branching.
What Changed
Context Snapshots
Agents now capture and restore their working state. When an agent is deep into a multi-step task and something fails, we were losing context. Now we snapshot the agent's state at key points—files written, API calls made, intermediate results—so we can resume or retry without starting from scratch.
This is implemented in context_snapshot.py and hooks into the task runner. The snapshots are lightweight JSON blobs stored alongside mission state.
Mission Grouping
We added the ability to group related tasks under a single mission. Previously, every agent task was atomic. Now you can define a mission with multiple subtasks that share context and dependencies. This makes complex workflows—like 'migrate this API, update the docs, notify the team'—explicit and trackable.
The executor now understands mission hierarchies and can parallelize independent subtasks while respecting dependencies.
Speculative Branching
This one's experimental. The speculative aggregator lets agents explore multiple execution paths in parallel and merge results. Think of it like Git branches but for agent reasoning. An agent can try two different approaches to solving a problem, compare the outcomes, and choose the best one—or combine them.
Right now we're using it for code generation tasks where there's no single 'correct' implementation. We let the agent generate multiple candidates and use heuristics (test coverage, complexity, performance) to pick the winner.
Why It Matters
These features move us from 'fire and forget' agent tasks to durable, resumable workflows. Context snapshots mean we waste less compute on retries. Mission grouping gives us better observability into what agents are doing and why. Speculative branching opens the door to agents that can evaluate trade-offs instead of committing to the first solution they find.
The changes touch core execution paths—agent_executor.py and task_runner.py—so we're rolling this out carefully. Early testing shows context snapshots reduce failed task retries by about 40%, which is meaningful at scale.
What's Next
We're watching how speculative branching performs in production. If it proves useful, we'll expand it to more task types and add smarter merging strategies. We also want to surface mission groups in the UI so you can see the full workflow tree, not just individual tasks.
Context snapshots need compression—right now they're verbose. We're exploring diffing strategies to store only deltas between snapshots.
And as always, we're listening. If you're using the agent executor and these features help (or cause problems), let us know. We ship fast and iterate.