The research agent that lives inside the Ditto web app — the one you talk to from the overview chat — has been rebuilt from the foundations up. The visible changes are speed, reliability, and an agent that actually knows what the API can do. The architecture underneath has changed too; this note covers both, briefly.
What you'll notice
Replies stream. The agent's response now appears word by word as it's generated, rather than landing as a single block after a pause. For long answers — synthesis across studies, multi-step recruitment plans — the difference between waiting eight seconds for a wall of text and watching the answer build in real time is the difference between trusting the agent and reaching for the refresh button.
Tool calls are far more reliable. When you ask the agent for something concrete — "what was the headline finding from the price-sensitivity study last month?" — it now executes the relevant tool call before answering, rather than answering from memory and getting the number wrong. Internally this is tool_choice="required" for retrieval-style prompts, but the customer-visible effect is that the agent stops making things up.
The agent now reads its own documentation. Every turn, the agent receives a contextual block called the Ditto MCP playbook — a versioned set of tool contracts, workflow recipes, and an error playbook covering the most common failure modes. The result: the agent knows about the new endpoints we shipped last week, knows that state is a 2-letter code rather than a full name, knows the difference between request_id and job_id, and so on. We update the playbook; the agent picks up the change on the next turn without a redeploy.
Progress updates during long work. When the agent is doing something that takes more than a moment — recruiting a panel, polling jobs, running synthesis — it now posts short status messages in the chat as it works ("recruiting panel… 7 of 12 confirmed"; "polling jobs… 4 of 10 finished"). No more silent waits.
Conversation replay. If you refresh the page mid-turn, the agent picks up exactly where it left off using an Idempotency-Key rather than restarting. SSE replay reconstructs the in-flight stream from the audited event log.
Under the hood
For the technically curious: this is a new microservice (agent_service) with FastAPI / SSE on the front and a clean separation between the chat orchestration layer and the persistence layer. Conversations, messages, runs, and run events are all audited to Postgres, which is what makes replay possible. Tools and credentials are resolved per-organisation per-request from a typed tool library; request-level overrides are constrained to a subset of the library allowlist so a misbehaving caller can't escalate.
We've moved from the old in-process LLM call site to OpenAI's Responses API, with MCP tool orchestration on top. None of those acronyms matter to the user; the visible effect is the speed and reliability story above.
What didn't change
The conversational surface — where you click to chat, what the agent looks like, the way you address it — is unchanged. The agent's name, its persona, and the URLs you talk to it from are all the same. Existing API key scopes are unchanged. The Slack agent (upgraded last week) shares the same underlying runtime improvements.
What this enables next
Two things, both due in the coming weeks:
One-off questions to a single persona or group, without the study scaffolding — landing this Friday.
Image and PDF attachments on questions, so the agent can show personas a logo or a one-pager and ask for reactions — also landing this week.
Neither is dependent on this rebuild, strictly, but both are easier to build well on top of the new runtime.
One operational note
Failed tool calls are now far more verbose in the chat — when something genuinely goes wrong, you'll see the underlying error rather than a generic "agent failure" message. Most of the time this is helpful; occasionally it's noisier than the previous behaviour. We're tuning the line between "useful detail" and "wall of stack trace" over the next couple of releases. Feedback welcome.
Full reference for the new agent service is in the API docs.
---