Over the past several years, the authors have been developing an instructional technology, called Teachable Agents (TA), which draws on the social metaphor of teaching to help students learn. Students teach a computer character, their "agent," by creating a concept map of nodes connected by qualitative causal links. The authors hypothesize that the metaphor of teaching allows students to use a well-known, and potentially productive, schema for organizing their interactions and for interpreting feedback (for reviews of the benefits of learning-by-teaching, see Annis, 1983; Biswas et al., 2005; Renkl, 1995; Roscoe & Chi, 2008). The authors have also conducted studies to examine the nature of the learning benefits from TA. Their studies on the effectiveness on TA have been, in the past, of relatively short duration, used specially designed content, and were taught under the strict edicts of the research designs. The authors wanted to see how the technology would fare in the more complex ecology of regular instruction, when teachers could integrate TA as they chose, into the flow of their normal curriculum over a sustained period of time. Of particular interest were 1) whether they could replicate their earlier PFL (Preparation for Future Learning) results showing that learning benefits persisted for students, "even when no longer supported by the technology"; 2) whether gains would emerge on the standard, basic-value assessments, in addition to their own added-value assessments that focused on the TA's strength of promoting causal reasoning; and 3) whether the learning benefits would be associated with particular student behaviors in the TA system. A small, local school district agreed to use the TA technology as added-value instruction to complement their regular science curriculum. The district had adopted the Full Option Science System (FOSS), developed by the Lawrence Hall of Science (www.lhsfoss.org). FOSS kits come complete with teacher guides, textbooks, videos, hands-on activities, worksheets, and assessments. The study involved six teachers and 134 5th-grade students (104 with permission to analyze their data). Teachable Agents weathered its first test in the complexity of the real world, where teachers chose how to use the software for several months as an added-value to their normal instruction. Students exhibited a deeper causal understanding of the FOSS material, as measured by the added-value tests and the "Why" questions in FOSS's own basic-value assessments. The TA activities did not displace basic learning from the FOSS kit. Moreover, the degree to which students used the map editing and feedback features correlated with learning, even after controlling for prior achievement. And finally, perhaps the most exciting result is that experience with TAs supported students' future learning of new content, even when they no longer used the software. (Contains 5 figures.)