Main finding
Splitting work across specialist agents with shared memory improves coverage and consistency, but self-correction only works when the correction comes from an external check. Agents that review their own reasoning tend to keep their original errors.
What supports it
- Simulated agents with memory, reflection and planning sustain coherent multi-step behaviour (Park et al. 2023).
- Verbal reflection improves later attempts when a feedback signal exists to reflect on (Shinn et al. 2023).
- Independent agents arguing to a conclusion reduce factual errors in some tasks, at a real cost in tokens and latency (Du et al. 2023).
- Without external feedback, models struggle to correct their own reasoning, which is why the verification step must be independent (Huang et al. 2023).
What this does not claim
- Errors compound across agents, and every additional specialist adds cost and latency.
- Shared memory helps consistency but does not guarantee that the shared fact is correct.
Planning assumptions used
- A named human owns the final publication of any agent output that reaches a client.
- Memory is auditable: every retained fact records its source and timestamp.
Open before signature
- What is the verification source of truth for each specialist?
- How long is memory retained, and how is it corrected when a source is wrong?
- What is the cost per completed analysis compared with the current process?
Sources
Figures marked illustrative are planning assumptions. This note separates published evidence from project assumptions and does not present a supplier quotation, a valuation or a forecast of returns.