Time-travel debugging for AI agents
Replay any agent run for $0. Fork any step. Prove the fix.
Flashback captures every model and tool call your agents make, from APIs and databases to browsers and sub-agents. Change a prompt, model or tool, prove it on past runs, and audit exactly what each agent did and why.
- Deterministic
- 100%
- Every captured step replays exactly as it was recorded.
- Replay cost
- $0
- Replaying a captured run calls no model and spends no tokens.
- Side effects
- 0
- Replay serves captured results, so nothing is charged, sent or written again.
- Action on record
- Every
- Each input, output and decision is kept, so you can show what an agent did and why.
Try it yourself
Traces show what happened. Flashback lets you run it again.
Reproduce, fix, prove: the debug loop every developer knows, now for agents. Click through one example run, a support agent that refunded a final-sale order, or let it play.
Captured runslast hour
all agents1 flagged1 error- RunsRuns captured
- 6
- ModelModel calls
- 12
- ToolsTool results
- 13
- ReplayReplay cost
- $0.00
Every model call and tool result is captured as the agent runs. Nothing to configure per run.
A support agent refunded a final-sale order. Open that run.
Testing against many past runs is in progress. Team features arrive with Flashback Cloud.
How it works
A few lines to capture everything.
Wrap your model and tool calls once. Flashback records each run from inside your code, in production or anywhere else. No new framework, no rewrite.
- Captures every model and tool call, with inputs, outputs, errors and timing
- Replays any run offline for $0, with no live side effects
- Forks from any step to try a new prompt, model, tool or code
- Keeps earlier steps frozen; everything after the fork runs live
- Inspects any run and previews forks from your terminal
- Any model provider
- Any agent framework, or none
Runs inside your code · No proxy, no network hop · Nothing leaves your machine unless you choose
from flashback import create_recorder, create_replayer, create_fork# one per runrec = create_recorder(agent_name="support-agent")# one per run# every model callllm = rec.llm(my_llm)# every model call# every tool calltool = rec.tool("lookup_order", lookup)# every tool call# run your agent as is …# saved as a cassettecassette = await rec.end()# saved as a cassette# replay for $0, no live callsreplay = await create_replayer(cassette["runId"])# replay for $0, no live calls# fork, earlier steps frozenfork = await create_fork(cassette["runId"], at=3)# fork, earlier steps frozenExamples
Built for agents that take real actions.
Debug what broke, explain what happened, and improve what already works.
Where it's going
Every run you capture becomes a test your agents must pass.
Today Flashback runs on your own machine: capture, replay and fork. Next, Flashback Cloud turns your team's captured runs into a shared library, a check on every pull request, and a record of what your agents did and why.
On your machine
- Capture every model and tool call, with inputs, outputs, errors and timing
- Replay any run offline, for $0
- Inspect each step's input and output
- Fork from any step, earlier steps frozen
- List, inspect and preview forks from the terminal
Test at scale
- Test one change against many past runs at once
- See which results changed before you ship
- Safe forks: after the fork point, real-world actions are held back by default and each step is marked frozen or live
Flashback Cloud
- Agents that fix agents: your coding agent replays the failure, writes a fix and tests it on past runs before you review
- A check on every pull request that replays past runs
- A shared run library: pull any captured run and replay it on your machine or in the browser
- Replay and fork in the browser, no local setup
- Audit export for compliance reviews
- Capture
- Replay
- Fork
- Proven fix
Stop guessing why your agents fail.
Flashback is in private early access. Start with one agent today; team library and CI checks follow with Flashback Cloud.