Agent Patterns
Patterns in action
Three deployment patterns that work in practice, built on the primitives the platform ships with: sandboxes, orchestration, and MCP connectors.
Autonomous coding pipelines
Agent squads take a ticket from triage to a reviewed pull request: exploration, code changes, tests, and a human sign-off gate.
Each agent runs in an ephemeral sandbox. Output lands in your repository as a normal branch — nothing executes outside the sandbox without your approval.
Explore the underlying capability$ synth agent squad --pipeline pr-review
▸ explore → reads repo, maps call sites
▸ plan → proposes 4 changes
▸ code → edits in sandbox
▸ test → runs suite, 14/14 pass
✓ branch ready for human review
Multi-step research workflows
A research agent plans a question tree, fans out to parallel queries, and synthesizes sources into a single cited report.
The planner owns the outline; worker agents own the sources. MCP connectors provide typed access to your data sources.
Explore the underlying capability$ synth agent plan --topic "mcp adoption"
▸ planner → 6 research questions
▸ workers → 6 parallel queries
▸ sources → 23 references collected
✓ report drafted, citations inline
Data extraction at your desk
Agents read documents, normalize fields, and write structured records back to your tools — through your own MCP servers.
Nothing is uploaded for training. Every connector runs against the data you grant it, scoped per task.
Explore the underlying capability$ synth agent extract --source invoices/
▸ read → 212 documents
▸ normalize → 14 fields per record
▸ write → MCP server: postgres.invoices
✓ 212 rows committed, 0 key leaks
Your workflow, your pattern
Bring your own keys, connect your own data sources, and compose the pattern that fits your team.